DK Street Journal

The companies that physically build AI infrastructure — the switch chassis, the full server racks, the busways and power modules hyperscalers bolt into data centers — spent late July reporting accelerating growth and raising guidance. Over four sessions from 17 to 21 August, the six largest listed ones fell double digits together, an average of 14.3%.

The trigger was not in their order books. On 18 August the 30-year Treasury yield touched 5.323%, its highest since 2007, on stalled talks to end the war with Iran, persistent inflation and uncertainty over policy under new Federal Reserve chair Kevin Warsh. That day Celestica fell 8.8%, Jabil 8.9%, Sanmina 8.5%, Benchmark 8.1%, Plexus 8.0% and Flex 7.6% — a range so narrow it names its own cause, since none of the six released anything. The following session AI cloud infrastructure names sold off again as a tally of roughly $3 trillion in off-balance-sheet AI commitments across nine large technology companies circulated. Long yields set the discount rate on earnings that sit in 2027 order books, and they price data-center project finance.

The control that wasn't

The cleanest test is Benchmark Electronics, a Tempe, Arizona builder of circuit boards and integrated systems for aerospace, defense, medical and semiconductor-equipment customers — the name here with the least compute exposure. It fell 13.3% across those four sessions, matching Celestica's 13.5%, and is down 14.5% over 30 days against Celestica's 13.2%. Whatever repriced, it was not AI program risk. Benchmark is not a pure control in any case: its advanced computing and communications line grew 71% year over year last quarter on an AI ramp, while aerospace and defense fell 12% on program timing. Group revenue rose 17.7% to $756m and the full year was raised to a record $3.0bn.

Celestica: the business went one way, the multiple the other

Celestica, the Toronto contract manufacturer that has moved up the value chain into designing its own 800-gigabit and 1.6-terabit Ethernet switches and full AI racks, is the sharpest divergence. Revenue growth has accelerated four straight quarters — 26.4%, 45.8%, 52.8% and 62.4% — to $4.699bn. Its Connectivity and Cloud Solutions arm grew 84% to $3.81bn with segment margin up to 8.7%, while the Advanced Technology Solutions side, serving aerospace, industrial and health customers, grew 8%. Hardware Platform Solutions, the own-design line, reached about $1.9bn. Operating margin widened to 9.75%. On 28 July management raised 2026 guidance to $20.5bn of revenue and $11.30 of adjusted earnings per share, and named custom racks for OpenAI's Jalapeno accelerator and a design role on AMD's Helios interconnect as 2027 programs.

That matters because hyperscalers increasingly bypass traditional server brands and contract directly with builders like Celestica, Foxconn and Quanta for whole racks — value capture moving upstream from the OEM to the assembler that owns the design.

The shares are flat over six months and down 16.9% over three. Because pass-through memory and component inflation — server memory contract prices rose 90–95% in one quarter — bloats the revenue line of a thin-margin conversion business at near-zero profit, price against gross profit is the honest lens. Celestica trades at 18.84x trailing gross profit and 14.07x forward, against 26.26x in mid-May.

The reasons for hesitation are about cash, not demand. Three customers were 32%, 17% and 14% of revenue, roughly 63% combined. Inventory reached $3.4bn, up $1.5bn year over year. Capital spending is guided at $1bn this year with a $1.5bn placeholder for 2027, against a trailing free-cash-flow yield of 1.52%. And on 6 August the company priced 9,677,419 shares at $310, about 15% below the prior close, raising $3bn and diluting holders by roughly 8.4%.

Flex: the margin sits in the part being spun out

Flex, the Singapore-founded manufacturer whose products run from automotive modules and medical devices to switchgear, busway and power distribution units, has also accelerated four quarters running, from 4.0% growth to 20.6% on $7.93bn of June-quarter revenue. Its Cloud and Power Infrastructure segment did $2.2bn, up 35%, at a 9.7% operating margin — well above the 6.7% adjusted corporate rate — with power up more than 70% and over 90% of the next three quarters booked. The other two segments are bigger and thinner: $3.1bn at 5.2% and $2.7bn at 6.6%. Flex guides to 23% revenue growth this fiscal year, and diluted shares have fallen from 400m to 374m. It has named the leadership for a spin-off of the cloud and power business in the first quarter of 2027. Flex trades at 14.70x trailing gross profit, down from 21.42x in May.

The other three followed rather than led. Jabil had gained about 16% from 21 July before surrendering all of it, leaving it down 1.85% over 30 days. Sanmina is the worst over three months at -20.6%. Plexus, the least AI-exposed by revenue mix, fell least at -6.5%. Benchmark, at 9.05x trailing gross profit with a 4.83% free-cash-flow yield, is the cheapest of the six on this measure.

One technical point does argumentative work: Celestica is the only one of the six whose 50-day average has fallen below its 200-day, and it crossed on 12 August — after the offering, before the yield spike. That is share-count and price mechanics, not results.

The setup

Where it stands — Six contract builders repriced together on a rates shock while five of six raised guidance. Would confirm — Celestica's Q3 revenue lands inside the $5.25–5.55bn guide with CCS margin at or above 8.7%. Would invalidate — A hyperscaler program slips, or Celestica's inventory rises again without matching revenue. Watch next — Celestica's third-quarter results in late October; Flex's spin-off targeted for the first quarter of 2027. Valuation — Celestica 18.84x trailing gross profit and 14.07x forward, against 26.26x in May; Flex 14.70x and 12.43x.

A consumable that got expensive

FormFactor sells probe cards: dense arrays of microscopic contacts that touch a chip while it is still on the wafer and test it electrically, before anyone spends money packaging it. For most of the past decade this was an unglamorous consumable business earning a gross margin in the high thirties. In the June quarter the company kept 50.7 cents of gross profit on each revenue dollar, against 37.2 cents a year earlier — four consecutive quarters of expansion totaling 13.5 points.

The mechanism is mix, not pricing power in the ordinary sense. Testing a stack of high-bandwidth memory (HBM) dies requires far more contacts, far tighter placement and data rates above 10 gigabits per second, and an advanced card of that kind sells for several times a legacy one. So a shift in what FormFactor ships lifts revenue per wafer tested without a matching rise in cost. DRAM probe-card revenue rose 48.9% year on year to $85m, with HBM about two-thirds of it, and two customers adopted a full-wafer contactor that tests hundreds of stacks at once.

Revenue grew 31.9% to $258.2m. Operating income grew 370%, taking operating margin to 22.4% from 6.3%. Management guided the September quarter to $270m and said the constraint on further share gain is its own capacity, not orders — a new Texas plant ramps from the fourth quarter through 2028, with capex of $140-170m this year. Its co-packaged-optics revenue, a second test market, was raised from $10-20m for the full year to more than $20m by the end of September. FormFactor competes for this work mainly with the Japanese card makers MJC and JEM; it has three HBM customers and strong position at two of them.

The four sessions that made the month

Measured to 17 August, the seven sub-fab and test suppliers grouped with FormFactor were up almost 3% on the month, with FormFactor itself up 22%. Then the bond market moved: the 30-year Treasury yield approached 5.33%, a two-decade high, and the Philadelphia Semiconductor Index fell more than 5% on 18 August. Over four sessions every one of these names fell between 13.8% and 20.8%. This rung of the supply chain books revenue against 2027 and 2028 capacity plans, which is precisely the duration a higher discount rate punishes hardest.

So the multiple compressed while the profits rose. FormFactor's price per dollar of trailing gross profit is 21.0x, down from 28.2x on 21 May and 23.7x in February, even as trailing gross profit rose 16% in one quarter. It trades at 37.5x forward earnings against 77.1x trailing, the gap being consensus for earnings of $3.05 this year against $1.13 last.

Where the numbers agree with the selling

Ichor Holdings, which builds the gas and chemical delivery panels that meter reagents into etch and deposition chambers for equipment makers, is the exception. It grew revenue 22.7% to $294.8m and raised full-year guidance to at least 30% growth. But it keeps 13.9 cents of gross profit per revenue dollar — against MKS's 47.7 — turned that into $1.0m of net income, and burned $15.9m of operating cash building inventory. It sold $200m of stock at $80.70 a share, 39% above where it now trades. Lam Research and Applied Materials together were 76% of 2025 sales. At 19.7x trailing gross profit it sits exactly where it did in February, and at more than twice the multiple of Ultra Clean Holdings, which does comparable made-to-print work for the same buyers.

MKS Instruments — vacuum, flow control, radio-frequency power and, since the Atotech deal, plating chemistry — accelerated to 28.3% revenue growth from 15.2% in March, guided September semiconductor revenue to better than 50% growth, and cut net leverage to 3.0x from 4.0x on $188m of quarterly free cash flow. Its price per dollar of gross profit fell from 11.9x in May to 9.8x.

The supposed cyclical laggards held up better than the artificial-intelligence names, which inverts the usual framing. Cohu, a back-end test-handler maker long described as automotive-levered, saw automotive revenue fall 24% while computing orders rose 150% to 46% of system orders; it raised full-year growth guidance to about 35%. Axcelis, which makes ion implanters, is the genuine laggard: gross margin fell to 42.4% from 44.9% and operating income dropped 30% on 10.6% revenue growth. Its forward price-to-gross-profit, 10.7x, sits above its trailing 10.4x — consensus expects the gross profit line to shrink.

The demand behind all of it is memory capital spending rather than logic wafer starts: SEMI forecasts DRAM equipment spending up 29% to $37bn this year.

The setup

Where it stands — FormFactor's margin and HBM mix improved through the June quarter while its multiple against gross profit fell roughly a quarter since May.

Would confirm — September-quarter revenue at or above the $270m guide with GAAP gross margin holding above 50%.

Would invalidate — DRAM probe-card revenue falling below $85m as mix reverts toward commodity DDR cards.

Watch next — FormFactor's third-quarter report in late October, and whether co-packaged-optics revenue passed $20m.

Valuation — 21.0x trailing gross profit and 18.9x forward, against 28.2x trailing in May and 23.7x in February.

Two companies with no product inside an artificial-intelligence server rack led the market's 800-volt power theme over the past month. Neither sells a converter, a magnetic or a power stage into a graphics-processor tray. One of them is barely a business yet.

Ideal Power, an Austin-based developer of a bidirectional solid-state switch it calls B-TRAN, recorded revenue of $5,800 in the second quarter against a net loss of $3.4m. It employs 17 people. The shares rose 17.3% over the past month, which values the company at roughly 2,660 times trailing sales — a number that means nothing except that revenue is not the thing being priced. What is being priced is a $400m sales "pipeline", split about evenly between automotive and data centers, and a co-developed intelligent circuit breaker for an unnamed US hyperscaler aimed at Nvidia's Rubin Ultra generation. Prototypes are due at the end of the fourth quarter of 2026. The company raised $27.7m in May, holds $41.3m of cash and guides to burning about $10.4m this year.

Ultralife, the month's best performer at 28%, makes lithium batteries and military communications gear and carries a market value near $118m. Its June-quarter revenue fell 1.3% to $47.9m. Gross margin rose 590 basis points to 28.9%, but 230 of those points came from a one-off refund of tariffs collected under the International Emergency Economic Powers Act (IEEPA). The genuine news was order flow: backlog reached a record $117.5m, up 39%, and roughly $130m by early August on defense pull-through. At 2.65 times trailing gross profit, 0.89 times book and 8.3 times forward earnings, the market is plainly not paying for artificial-intelligence content here. It is a defense restocking story that happens to sit in the same neighborhood.

The suppliers that ship fell instead

Vicor, which makes the modular direct-current converters that step a rack bus down to the volt-level rails a processor draws, had the best quarter of its cycle. Stripping the prior year's $45m patent settlement, core revenue grew 49.3% to $143.4m. Gross margin expanded 280 basis points sequentially to 58.0%, far above the 52.6% it earned across 2025. One-year backlog rose 145% to $379.7m, and full-year guidance went above $600m. The shares fell 24.9% over three months. Price per dollar of trailing gross profit has compressed from about 43.8 times in May to 33.9 times — still nearly four times the 9.1 times of a year ago. Its litigation is live: the US International Trade Commission opened a new Section 337 investigation in February naming Monolithic Power Systems, the rival that took the H100 socket Vicor lost.

Bel Fuse, a maker of magnetics, alternating-to-direct-current power supplies and Cinch connectors, is the sharpest disconnect. Revenue growth accelerated to 25.2%, gross margin rose 120 basis points and operating income grew 47.4% — roughly twice the pace of sales. Data solutions revenue reached $58m, up 55%, on the start of a high-performance-computing ramp. Defense, at $66.5m, is still the bigger line. The stock fell 8.8% in the month after that beat-and-raise, leaving 26.0 times forward earnings against consensus for 45% earnings growth this year.

Navitas is the one name whose decline the numbers support. The gallium-nitride and silicon-carbide designer is one of ten silicon suppliers named in Nvidia's 800-volt reference architecture, and by far the smallest. June-quarter revenue was $10.5m, down 27.3%. Across four quarters it has produced minus $0.6m of gross profit on $36.5m of revenue. Diluted shares are up 18.6% year on year, including $373m raised at $21.89 — some 40% above where the stock now trades. Even after a 46.5% three-month fall it is held at 92.7 times trailing sales.

Why good quarters did not pay

The gating fact is timing. Full-scale 800-volt deployment arrives with Nvidia's Kyber racks in 2027; Navitas dates its sidecar racks to mid-2027. That is five or six quarters of waiting, and long-dated revenue is what a discount rate punishes. In late July the PHLX Semiconductor Index shed more than $1trn over three sessions as 30-year Treasury yields touched 19-year highs. Vicor and Navitas both slipped into downtrends by 20 August, their 50-day averages crossing below their 200-day, after trading in clear uptrends in May. Bel Fuse's uptrend weakened over the same weeks.

So the theme did not stall. It rotated — out of the companies reporting the revenue and into the two that have none to report.

The setup

Where it stands — Ultralife and Ideal Power carried the group's month while Vicor, Bel Fuse and Navitas fell despite two beat-and-raise quarters. Would confirm — Bel Fuse third-quarter sales landing at or above the $205-225m guided range with gross margin holding 39-41%. Would invalidate — Vicor's one-year backlog falling below $379.7m, or book-to-bill dropping under 1.0. Watch next — Ideal Power's hyperscaler circuit-breaker prototypes, due at the end of the fourth quarter of 2026. Valuation — Vicor is at 33.9x trailing gross profit and 26.6x forward, against 43.8x in May and 9.1x a year ago.

AvePoint sells the control layer between Microsoft 365 and the data inside it: backup, permissions, retention policy, and — since July — a standalone product that inventories and polices the artificial-intelligence agents an enterprise has switched on. On 6 August the company reported that annual recurring revenue (ARR) grew 27% to $465.1m, with record net new ARR of $29.9m, up 35%. That was the thirteenth straight quarter of double-digit growth in net new ARR. The shares fell 4.7% that day.

They have not recovered since. AvePoint closed 21 August at $13.08, within half a percent of where it began the month, while UiPath — the low-code robotic-process-automation (RPA) vendor whose robots execute rule-based workflows for banks, insurers and governments — rose 32% over the same stretch on no company disclosure at all.

The one with the numbers

AvePoint's case is that nobody turns on Microsoft's Copilot, or an agent, over ungoverned data. Management says its AgentPulse control packages roughly doubled quarter over quarter at deal sizes two to three times larger, that control products account for about 40% of pipeline, and that some customers already manage more than 5,000 agents, with counts doubling every quarter. Revenue grew 22% to $124.5m; software-as-a-service revenue reached 79% of the total. Full-year ARR guidance went up to $522.1-528.1m.

The competitive question is unavoidable: Microsoft shipped its own Agent 365 on 1 May at $15 per user standalone. AvePoint's answer is that it governs Google Workspace, Salesforce, Atlassian and Okta as well, and that it reaches customers through managed service providers — now 59% of ARR, up from 56%, and the source of 67% of incremental ARR. Customers above $100,000 of ARR numbered 911, up 26%, the best growth in three years. Gross margin slipped 94 basis points to 73.1% on mix, while GAAP operating margin widened 130 basis points to 8.2%.

The one with the story

UiPath's April-quarter revenue accelerated to 17.3% growth, but ARR grew only 12%, to $1.901bn, with dollar-based net retention of 109%. Management's own guidance puts July-quarter revenue at $395-400m — growth of 9.2% to 10.6% — and full-year ARR at $2.051-2.056bn, implying deceleration to about 10.8%. Gross margin fell 97 basis points to 81.1%.

On 20 August the company introduced Maestro Flow, an orchestration canvas that runs inside Claude Code, Cursor, GitHub Copilot and VS Code — a claim to sit above the coding agents rather than be replaced by them. It arrived after most of the share-price move, with no pricing and no revenue line. At $15.90 the stock sits roughly 19% above the $13.25 average analyst target, with sixteen of twenty ratings at hold.

Measured per dollar of trailing gross profit — the only lens that survives margins ranging from 70% to 83% across these businesses — UiPath now trades at 6.28x, against 4.19x three months ago and 4.45x in February. The multiple expanded about 50% while trailing gross profit grew 3.5%. AvePoint trades at 8.28x against roughly 12.3x a year ago, with trailing gross profit up 22% over that year. The business went one way; the price went the other.

The control fails

If this were an automation re-rating, the members without agent revenue would have lagged. They did not. SPS Commerce, which runs the electronic trading network connecting suppliers to retailers, grew revenue just 5.6% and expanded its price-to-gross-profit multiple 43% in three months, to 5.41x — nearly matching UiPath. Its gain is dated and specific: an 11.5% session on 31 July after second-quarter earnings of $1.27 a share beat $1.12, with adjusted EBITDA up 19% and trailing free cash flow up 40% to $198.7m.

CCC Intelligent Solutions, whose network links collision repairers to property-and-casualty insurers, rose 22% on takeover arithmetic: Bloomberg reported on 18 August that Copart is in talks to buy it, against private-equity bidders, with Morgan Stanley running the process. EverCommerce fell 15.9%, growing revenue 2.7% at 94% net revenue retention, with management blaming AI-driven search behavior for weaker organic customer acquisition — the clearest hard evidence of AI damage anywhere in the group.

The common cause is broader. Enterprise software has been re-rating since ServiceNow's 22 July results reopened the argument over whether AI kills seat-based software, amplified by forced unwinding from a collapsed leveraged fund, and layered on the best first half for US small caps since 1991. What lifted these five was their size and their sector, not their agents.

The setup

Where it stands — The group's best fundamental result, AvePoint's, produced no share-price gain; the biggest gain, UiPath's, rests on no new disclosure. Would confirm — UiPath's 3 September print showing ARR growth at or below the guided ~10.8% while the multiple stays above 6x. Would invalidate — UiPath disclosing a quantified agentic revenue line and lifting full-year ARR guidance above $2.056bn. Watch next — UiPath reports fiscal second-quarter results after the close on 3 September 2026. Valuation — AvePoint 8.28x trailing and 7.56x forward gross profit versus 12.3x a year ago; UiPath 6.28x and 5.91x versus 4.19x in May.

The pure-play that wasn't thin

Super Micro Computer, the San Jose builder of liquid- and air-cooled server racks that ships full AI clusters to data centers, told investors its June-quarter gross margin would land between 8.2% and 8.4%. It came in at 17.5%, which management attributed mainly to customer and product mix. Revenue nearly doubled, to $11.12bn. Operating income reached $1.49bn, a 13.4% operating margin — higher than either of the two far larger incumbents managed in their own latest quarters.

That matters more than the beat itself. The durable argument against server original equipment manufacturers (OEMs) is that they buy the expensive parts — accelerators from NVIDIA, memory from three suppliers — and add assembly that anyone can add, so margin gets competed to nothing as volumes rise. A company with about 6,200 employees, guiding fiscal 2027 revenue to $65–72bn against roughly $54.4bn of analyst consensus after taking more than $60bn of orders in a single quarter, is not the picture that argument predicts. Management did caution that some of those commitments are not firm and can shrink before they become revenue.

Dell's margin fell where it didn't count

Dell Technologies, the Round Rock maker of PowerEdge servers, storage and PCs and the largest AI-server vendor by share, saw April-quarter revenue rise 87.5% to $43.84bn. Gross margin compressed 337 basis points, to 17.75%. The mechanism is memory: DRAM contract prices rose 90–95% in that quarter and again by more than half in the next, because the three memory makers are diverting capacity to the high-bandwidth memory that AI accelerators consume.

The compression did not reach the profit line. Dell's Infrastructure Solutions Group posted record revenue of $29bn, up 181%, with segment operating income up 206% to $3.1bn — income outgrowing revenue means segment margin expanded, to roughly 10.7%. Consolidated operating margin went to 8.34% from 5.27%. The company booked $24.4bn of AI orders, exited with a $51.3bn backlog and raised its fiscal 2027 AI-server revenue goal to $60bn.

HPE bought its way up the margin stack

Hewlett Packard Enterprise sells ProLiant servers, storage and — since the Juniper acquisition — enterprise and data-center networking. Its gross margin moved the opposite way from Dell's, to 36.5% from 27.6%, because networking revenue grew 148% to $2.7bn while the server line grew 33% to $5.5bn. Mix, not pricing, did that. Net leverage is down to 2.3x and CFO Marie Myers said the 2.0x target arrives a year early, with full-year free cash flow guided to at least $3.5bn.

Two cautions sit against that. Consensus has HPE decelerating to roughly 12% revenue growth next fiscal year, so this is a step, not a run-rate. And AI networking is arriving margin-dilutive for the incumbents Juniper is attacking: Cisco guided gross margin down toward 64.5%, and Arista's management has said it has no pricing power until 2027. On the server socket itself, Scott Black told the Barron's Midyear Roundtable that Dell's hyperscaler machines beat HPE's "hands down" on total cost of ownership.

Same demand, three prices

Here the businesses and the shares part company. Dell's price per dollar of trailing gross profit went from 3.69x in February to about 11.5x now, while those gross-profit dollars grew 19% — the multiple did roughly nine-tenths of the work. Forward earnings went from about 18x in May to 23.5x. HPE re-rated from roughly 2.7x trailing gross profit to 5.53x on 30% gross-profit growth, and from the 10–13x forward it carried in May to 15.6x.

Super Micro, with the same demand and a better operating margin, trades at 8.58x forward earnings and 3.32x forward gross profit — under half Dell's 8.90x. It is down 12.9% over twelve months while Dell is up 240%. The discount is not free: trailing free cash flow yield is minus 28.9%, against HPE's 5.6% and Dell's 3.2%, so the backlog is being funded by consuming cash, and the 2025 accounting allegations and March 2026 indictments of associates still sit on the file.

All three have been in uptrends only briefly together — Dell since late March, HPE since late April, Super Micro only since 19 August, after its results. Both incumbents have cracked: Dell is 11.3% below its 14 August peak and HPE 9.6%, on pre-earnings de-risking and a Netlist complaint at the International Trade Commission over DDR5 memory patents naming Supermicro, HPE and Lenovo. In mid-July the group round-tripped once already, Dell falling 14% in a session on positioning rather than news.

The setup

Where it stands — Integration margin is holding across all three builders, but only Super Micro's shares are priced as though it might not. Would confirm — Dell's ISG operating margin holds above 10% and backlog rises again from $51.3bn. Would invalidate — Super Micro's fiscal Q1 gross margin reverts toward its old 8–10% range, showing June was one-quarter mix. Watch next — Dell reports fiscal Q2 on 1 September; HPE reports fiscal Q3 on 2 September. Valuation — Dell 34.5x trailing and 23.5x forward earnings against ~18x in May; HPE 15.6x forward; Super Micro 8.6x.

Fervo Energy, a Houston driller that fractures hot rock the way shale producers fracture oil-bearing rock, told investors on 12 August what its first commercial power plant will actually earn. The answer, for 2027, was $60m to $80m. Management attributed the wide range to curtailments on the transmission lines leaving its Cape Station site in Utah, after the line's operator added another asset to the network — a queueing problem on somebody else's wires, not a failure of the rock. The shares fell 16.6% that day on a quarterly net loss of $55.9m, or -$0.38 a share against estimates nearer -$0.09.

That single session is the hinge of a month that looks, from a distance, like a broad flight from unbuilt electricity. It isn't.

The one with the contracts fell hardest

Three small companies now sell investors the same promise — firm power for artificial-intelligence data centers, delivered before the grid can. They are not the same business. Fervo owns 658 MW under binding power purchase agreements representing a $7.2bn revenue backlog, half of it to hyperscalers, and has mechanically completed two of the three 33-MW blocks that make up Cape Station Phase I, targeting first power this quarter. Oklo, the Santa Clara developer of 15-to-75 MW sodium-cooled Aurora fission plants backed by OpenAI's Sam Altman, has an order book near 14,100 MW built on a master agreement with data-center landlord Switch and letters of intent its own filings warn are contingent and non-binding. Fermi, a 35-employee Amarillo landowner classified as a specialty real-estate trust, is assembling a 4.8 GW power-and-data-center campus and had no binding customer at all until 10 August.

Since 22 July, Fervo is down 39.0%, Fermi 9.5% and Oklo 4.7%. Fervo alone accounts for roughly three-quarters of the decline in the three. The ranking on enterprise value per contracted megawatt runs precisely backwards to the price action: Fervo's $2.8bn of enterprise value against 658 MW is about $4.2m per contracted megawatt; Fermi's is roughly $17m; Oklo's contracted denominator is zero.

What each one actually published

Oklo's quarter was its best of the year. Its Groves isotope reactor in Texas reached first criticality inside eleven months of groundbreaking, the first privately funded reactor to do so on greenfield land, and the Idaho Aurora plant kept its 2028 start-up. It disclosed $3.0bn of cash and securities. It also raised 2026 operating cash use to $120–150m from $80–100m and capital spending to $400–500m, booked $1.21m of revenue against a $73.2m operating loss, and grew its diluted share count 25.8% year over year funding at-the-market issuance. Consensus has it earning $10.5m of revenue in 2027 and losing money every year through 2030. At 2,404x forward sales, the de-rating has arithmetic behind it.

Fermi went the other way on news: a 15-year binding lease with TensorWave covering 222 MW for about $6.5bn, plus a $431m convertible at 5.00%. That is under 5% of planned capacity, and the investment-grade guarantor behind it was still unnamed on the earnings call. Fermi held $91.7m of cash at 30 June; its own benchmarks of $3–4m per MW for power and $10–12m per MW for the data center imply $8–10bn to build the 640 MW it wants live by late 2027. Its share count rose 75.4% in a year.

Fervo, meanwhile, raised its 2030 capacity target to 1.1 GW from 1.0 GW on the same call, holds $2.1bn of cash — 43% of its market value — and trades at 1.79x book, the cheapest of the three on assets. Baird cut its target to $35 from $50 while keeping an Outperform rating; Jefferies and Bank of America upgraded to Buy on lower targets.

The revenue-bearing control refuses to cooperate

If this were a de-rating of nuclear-adjacent power generally, the companies with operating assets would have fallen too. Cameco, the Canadian uranium miner and fuel-services group, rose 10.3% over the same 30 days despite a quarter in which revenue fell 7.2% to $814m and net income dropped 92% on a missing Westinghouse payment. Explorer NexGen rose 12.3%, enricher Centrus 2.6%, reactor designer NuScale 7.3%.

The exception is BWX Technologies, sole supplier of naval nuclear propulsion components to the US Navy, which fell 10.2% while growing revenue 18.0% to $901.6m and lifting backlog 40% to $8.4bn. At 33.2x forward earnings it is the only anchored multiple in this article, and it fell anyway — a reminder that when long-bond yields hit a 19-year high on 18 August, with the 30-year at 5.34%, everything sold to the AI build-out trades as one basket. That move amplified rather than caused: Fervo was already down 23.1% before its earnings.

On the trend, Oklo's 50-day average has sat below its 200-day since 11 June and Fervo's crossed below on 29 July. Fermi's repaired to neutral on 28 July and has held. The company with the worst balance sheet has the best-behaved chart.

The setup

Where it stands — The only developer with binding contracted megawatts trades at the lowest price-to-book of the three, on guidance the market read as a downgrade. Would confirm — Cape Station Phase I delivers first power in Q4 2026 and 2027 revenue lands at or above $80m. Would invalidate — Further curtailment pushes 2027 revenue below $60m, or Fervo raises equity below book value. Watch next — Fermi must name the investment-grade guarantor behind the TensorWave lease; Oklo's Aurora-INL start-up remains dated 2028. Valuation — Fervo 1.79x book and 837x forward sales; Oklo 2.24x book and 2,404x forward sales; Fermi 3.69x book.

A guidance raise that counted for nothing

American Electric Power, the Columbus, Ohio company that generates and delivers electricity to five and a half million customers across 11 states, told investors on 30 July that large industrial load under contract through 2030 had reached 69 gigawatts. That was six gigawatts more than a quarter earlier. The same morning it raised 2026 operating earnings guidance to a range of $6.25 to $6.55 a share, from $6.15 to $6.45, and reaffirmed a five-year capital plan of $78bn — roughly double the plan of four years ago — implying 11% annual growth in rate base, the asset pile regulators allow it to earn a return on.

The stock fell 4.9% over the month to 21 August.

It was not alone, and the company it kept is the point.

The control group owns no data centers

Duke Energy, CMS Energy and DTE Energy each have junior subordinated notes listed in $25 denominations — "baby bonds," bought largely by retail investors for the coupon. They pay a fixed rate. They have no claim on a single megawatt of hyperscaler load. Over the same 30 days, Duke's fell 3.4%, CMS Energy's 3.8% and DTE's 3.9%.

Dominion Energy fell 3.9% over that stretch. NextEra fell 3.9%. American Electric Power fell 4.9%. The equities of the three utilities most exposed to artificial-intelligence power demand in America delivered, to within about a point, what a fixed coupon delivered. Behind it, the 30-year Treasury yield topped 5.33% on 18 August, its highest since June 2007, on a widening federal deficit and sticky inflation. Every utility is valued against that yield, because the dividend is the competing product.

Stretch the window and the picture inverts. Over twelve months Dominion returned 9.6%, American Electric Power 9.7% and NextEra 11.0%, while the three baby bonds lost between 9.8% and 11.4%. That is roughly 20 points of spread — the market's price for load growth. Over the last month, that spread was zero.

What actually happened at the businesses

Dominion, the regulated monopoly serving about 3.6m electric customers in Virginia and the Carolinas, sits on top of the densest concentration of data centers on earth in Loudoun County. Its contracted book across all stages reached 53.8 gigawatts by July, with roughly 12 gigawatts under signed electric service agreements — binding contracts, up from 10.4 gigawatts in May. Virginia segment operating earnings rose 22% to $670m.

Dominion's reported profit looks nothing like that: GAAP net income fell 55% to $340m on $704m of impairments, chiefly on non-regulated renewable natural gas plants. Its trailing price-to-earnings ratio therefore rose to 23.04x from 18.58x in May while the stock went nowhere — the multiple expanded because the earnings shrank. Construction risk is still live: the 2.6-gigawatt Coastal Virginia Offshore Wind project, 81% complete, saw its estimate rise nearly $300m to about $11.7bn.

Dominion is no longer an independent read in any case. On 15 May it agreed to an all-stock combination with NextEra at a fixed 0.8138 NextEra shares apiece. Dominion's discount to that ratio has narrowed from 6.8% in May to 2.3%. Its relative firmness is arbitrage convergence, not Virginia.

NextEra — Florida Power & Light plus the largest US developer of contracted clean generation — grew second-quarter revenue 12.4% to $7.53bn, added 3.6 gigawatts to a 35.1-gigawatt backlog, and raised its expectation for large load at the Florida utility to 8 gigawatts by 2032 from 6. Each gigawatt is worth roughly $2bn of capital spending earning a regulated return. Erste Group nonetheless cut it to Hold in June on financing costs, not demand.

The one fundamental crack

American Electric Power owns close to 90% of the nation's 765-kilovolt transmission — the highest-voltage backbone, and the reason hyperscalers negotiate with it rather than around it. It has collected $2bn of cash and collateral from customers backing 45 gigawatts of Texas load, which is real money against non-binding intent.

That 45 gigawatts is now frozen. On 3 August Governor Greg Abbott ordered an audit of data-center projects before further interconnections advance, and ERCOT suspended the classification notices due on 7 August, with the review expected to take months. The grid operator is fielding some 474 gigawatts of requests, about 90% of them data centers — more than five times its record peak demand. AEP's $78bn plan assumed only 13 gigawatts in Texas; scaling to 45 implies roughly three times the incremental Texas spending. Morgan Stanley and Truist trimmed price targets in mid-August.

At 18.96x forward earnings and 13.83x trailing enterprise value to EBITDA, AEP is the cheapest of the three — Dominion is 15.20x and NextEra 15.92x on that measure — against a targeted earnings growth rate above 9%. NextEra, on its own 2026 adjusted guidance of $3.92 to $4.02, prices near 21x for 8%-plus growth.

The setup

Where it stands — Load growth kept compounding through July; for one month the shares were paid a bond's return for it. Would confirm — Long yields ease while contracted load and capex guidance keep rising at the next quarterly reports. Would invalidate — AEP's contracted large-load figure falls below 69 gigawatts, or Texas customers withdraw collateral. Watch next — ERCOT's data-center audit conclusion, and NextEra and Dominion shareholder votes expected early September. Valuation — AEP at 20.78x trailing and 18.96x forward, versus 20.58x trailing in May; NextEra near 21x on guidance.

Three companies that carry other businesses' phone calls and text messages reported second-quarter results within a fortnight of one another this summer, and they went in three directions. RingCentral printed on 23 July and jumped a quarter in a session. Bandwidth printed on 29 July and lost nearly a third. Twilio printed on 6 August and rose 27% the next day. Whatever explains this group, it is not one rates trade or one short squeeze.

The unexpected winner is the one that was supposed to be structurally disadvantaged.

The seat-priced incumbent

RingCentral sells cloud phone, video and contact-center software to North American businesses on a per-seat subscription — the model that ought to be threatened if artificial-intelligence agents replace human staff at the desk. Its revenue grew 5.9%, to $657m, the slowest of the three by a wide margin. But it is the only one of the three that publishes a quantified AI figure: customers paying for at least one AI product now account for about 13% of its roughly $2.8bn of annual recurring revenue, twice the share a year earlier, with higher revenue per account and retention above 100%.

The rest of the print was mundane in the way that pays. Gross margin rose 73 basis points to 71.9%, because a subscription business carries no carrier fees in its cost of sales. GAAP operating income rose 36% to $50.3m. Management raised full-year revenue, margin and free-cash-flow guidance, lifted non-GAAP earnings guidance to $4.96–5.10 a share and raised the quarterly dividend by roughly two thirds to $0.125. Its balance sheet, long the bear case, is now unremarkable: $609m of convertible notes were repaid at maturity in March, net leverage is 1.5 times and nothing else matures until 2030.

Even after rising 77% in a month, the shares change hands at 3.1 times trailing gross profit, 13.3 times forward earnings, on an 11.6% trailing free-cash-flow yield. That multiple was 1.75 times the day before the results.

Where the metered model leaks

Twilio sells the voice and messaging plumbing that developers embed in their own applications, billed by usage — the purest way to own an AI agent's traffic. Its revenue growth accelerated for four straight quarters, from 14.3% to 22.0%, dollar-based net expansion reached 116%, and management raised full-year reported growth guidance to 18–18.5% from 14–15%. The business is working.

The leak is underneath. Messaging grew 28%, but only about 18% excluding US carrier pass-through fees, which were $71m in the quarter and are running toward $250m for the year. Those fees inflate revenue and sit in cost of sales, so trailing twelve-month gross profit rose 9.1% while the share price roughly doubled from $113.14 in February. Price to trailing gross profit went from about 6.9 times to 12.6 times over six months. And Twilio guided third-quarter organic growth down to 11–12% from 17%. It discloses no AI revenue line at all — its evidence is two unnamed conversational-AI customers that scaled to $6m and $9m of annual spend.

Bandwidth runs its own voice and messaging network out of Raleigh, North Carolina, which should make it the toll-taker rather than the toll-payer. Its 10-Q says otherwise: first-half cost of revenue rose $61m, of which $41m was higher pass-through messaging surcharges, taking gross margin to 36%, down four points. Reported growth of 22.2% became gross-profit growth of 9.6%, and the shares fell 29% on the day despite a raised outlook. That is a quality problem, not a demand problem — voice revenue grew 9%, software services 66%, and every one of five new $1m-plus wins included its orchestration layer or AI services, including production traffic from Salesforce's Agentforce contact center. Nor is debt the issue: $316m of zero-coupon notes due 2032 refinanced the 2028 converts, leaving net leverage at 1.6 times.

The mechanism behind both is the same. US carriers add roughly $0.003 to $0.005 per text message on top of the base rate, a markup of a third to more than half, while per-message prices themselves decline mid-single digits annually. Metered agent traffic arrives with a toll attached. Seat pricing does not.

What the charts say

Twilio has held a clean uptrend since 17 April, its 50-day average above its 200-day for some 84 sessions, and the business supports the direction if not the price paid. Bandwidth's trend broke on 6 August and the shares are down 39% from their 9 July high, at 4.96 times trailing gross profit against about 1.65 times six months ago. RingCentral, oddly, never established a trend at all: the gap-up on 24 July jumped straight over its own moving averages and the shares have shown no directional trend since 28 July.

The setup

Where it stands — The group's slowest grower carries the only disclosed AI revenue metric and the cheapest multiple; the fastest grower carries the highest. Would confirm — RingCentral's AI-paying share of recurring revenue rising above 13% in the third quarter, with gross margin still expanding. Would invalidate — Twilio's third-quarter organic growth landing below the guided 11–12%, or RingCentral's recurring revenue growth slipping under 5%. Watch next — Third-quarter results: RingCentral in late October, Bandwidth and Twilio in early November. Valuation — RingCentral 3.1x trailing and 3.0x forward gross profit, against Bandwidth's 4.96x and Twilio's 12.6x.

Cognizant Technology Solutions, the Teaneck, New Jersey outsourcer that runs applications, back-office operations and now AI deployments for banks, insurers and drugmakers, raised its full-year profit guidance on 29 July. A week later EPAM Systems, a digital-engineering firm that builds custom software for Western enterprises out of delivery centers across Europe, Latin America and India, cut its revenue forecast and lost 14.5% of its value in a session. Accenture, the Dublin-headquartered integrator with 799,000 staff and the largest consulting book in the world, said nothing at all — its last disclosure was 18 June.

By 21 August all three had risen roughly 10% in four sessions, and EPAM had closed the entire gap it opened on its downgrade.

Four sessions, one mechanism

The move had nothing to do with any of them. On 19 August the US Treasury said it would at least double the size of its debt buybacks, and long-end yields fell hard — the 30-year down more than 10 basis points to 5.184%. Nvidia fell 4.5% over those same four sessions and the S&P 500 slipped 1.0%. Cheap, slow-growing services businesses absorbed the money leaving expensive, fast-growing hardware.

It was the second such episode in a month. The first ran 22–29 July, when chip stocks shed more than $1trn and Accenture rose 23.6%. Between the two legs, across thirteen sessions, Accenture fell 0.8%. Its entire 34% month landed in nine trading days.

Infosys and Wipro, the two large Indian-listed outsourcers whose economics are most exposed to the rupee, gained 10.6% and 4.4% over the month — a quarter of Cognizant's 45.2%. Wipro actually fell 1.8% during the August leg. Whatever moved these shares, it was not offshore demand or currency.

Which of them earned it

Cognizant did. Second-quarter revenue reached $5.5bn, up 4.5%, with financial services up 12% and adjusted operating margin at 16.0% — a sixth consecutive quarter of expansion. Trailing bookings hit $29.1bn, including seven contracts worth $100m or more, and management raised adjusted earnings guidance while buying back $1.1bn of stock in the quarter. Management also disclosed something more interesting than the numbers: fixed-price and transaction-based work has grown as a share of mix for three straight years, and some clients now ask for rate cards with model training and inference costs embedded. That is outcome pricing arriving through the contract, not the headcount.

The catch is that Cognizant's trailing gross profit, at $6.93bn, is about 1% below the prior twelve months. The margin came from discipline and buybacks. Its price per dollar of gross profit has gone from 2.81x a month ago to 4.02x — the largest round trip in the group, and within 13% of where it stood a year ago.

EPAM did not earn it. Organic constant-currency growth was 3.4%, the Americas — 57% of revenue — grew 0.5%, and management blamed its own sales execution rather than the economy. Free cash flow was minus $18m against plus $43m a year earlier. What it does have is leverage: operating margin rose to 10.8% from 9.3%, with operating income up 20.4% on revenue up 4.5%.

Accenture is the ambiguous one. Third-quarter new bookings of $19.3bn were down from $19.7bn a year earlier, with full-year growth guided to 3–4% in local currency. Two things widely assumed about it are wrong. It stopped breaking out advanced-AI bookings after reporting about $11.5bn cumulatively across 11,000 projects, so the metric everyone cites cannot be tracked forward. And it is not shrinking: headcount rose to 799,000 from 779,000 at fiscal year-end. Revenue per employee is up about 5.7% to roughly $91,500 because revenue grew faster than hiring, not because anyone was replaced by an agent.

Who they are now competing with

The durable question is not whether agents delete billable hours. It is who sells the agents' installation. OpenAI has capitalised a deployment company at more than $4bn with TPG, Advent, Bain Capital and Brookfield, and Anthropic has a services joint venture reportedly valued above $1.5bn with Blackstone, Hellman & Friedman and Goldman Sachs. The model vendors now bid for integration work directly. Cognizant's response is to certify more than 10,000 engineers on Anthropic's Claude; EPAM has 5,700 and is courting OpenAI's partner network; Accenture joined OpenAI's alliance and built a forward-deployed engineering practice with Microsoft. All three are training staff on the platforms of firms that have just entered their business.

Accenture trades at 14.7 times trailing earnings and 13.4 times forward, against multiples of 21x to 37x at each of the past five fiscal year-ends, with a trailing free-cash-flow yield above 11%. Cognizant is at 10.8x forward, EPAM at 8.4x. Nine sessions of rotation restored Accenture's multiple only to where it sat in May, and left it a third below a year ago. The re-rating, such as it is, has barely begun — and so far only one of the three has produced a quarter to justify one.

The setup

Where it stands — Two rate-driven rotation legs lifted the sector 25% in a month; only Cognizant's business improved inside the window.

Would confirm — Accenture's late-September fiscal fourth quarter showing new bookings back above $20bn with book-to-bill above 1.0x.

Would invalidate — EPAM's next quarter holding organic constant-currency growth at 2–3% while free cash flow stays negative.

Watch next — Accenture reports fiscal Q4 and full-year 2026 in late September 2026.

Valuation — Accenture 14.7x trailing and 13.4x forward, versus 21x–37x at each of the last five fiscal year-ends; Cognizant 10.8x forward, EPAM 8.4x.