DK Street Journal

A backlog that quadrupled, and a share price that didn't

CoreWeave rents Nvidia graphics processing units (GPUs) by the hour and under multi-year contract to artificial-intelligence labs and large enterprises. It was incorporated in 2017 as a crypto-mining venture called Atlantic Crypto. In the June quarter it reported revenue of $2.575bn, more than double a year earlier, and disclosed a revenue backlog of $104.2bn, up 246% — a figure that excludes over $25bn of further commitments signed in the opening weeks of the third quarter. Live power reached 1.5 gigawatts after 500 megawatts were energised in three months, against 4.2GW contracted. Full-year revenue guidance went up to $12.4-13.2bn.

The shares are down 16.7% over three months and sit 52% below their June 2025 record.

That has done something specific to the price. In early May the market paid $16.40 for every dollar of CoreWeave's trailing gross profit. It now pays $9.37, and 5.51 times the gross profit expected on forward estimates. The multiple has halved while gross profit itself nearly doubled — the clearest divergence in the group between what the business reported and what the equity was marked at.

The bear case is properly funded

What the multiple is discounting sits below the gross-profit line. Gross margin before depreciation has compressed for four consecutive quarters, to 65.9% from 74.2% a year ago. Depreciation and amortisation of $1.393bn equalled 54% of revenue, which is most of the reported gross profit. Operating margin on generally accepted accounting principles was minus 1.9%, and the net loss widened to $626m.

The second charge is the financing. Interest expense reached $640m from $267m a year earlier, and management guided the September quarter to $860-940m. That is growth of well over 100% against revenue growth of 112% — the two lines are racing. Against roughly $30bn of debt, CoreWeave's senior notes carry coupons of 9.00% to 9.75%, and a recent $2.6bn delayed-draw loan priced at the Secured Overnight Financing Rate plus 550 basis points, a full point wider than May's facility, even after a landmark $8.5bn investment-grade-rated GPU-backed facility cut the blended cost. Capital expenditure guidance of $35-39bn is roughly three times this year's revenue. Every incremental megawatt is funded at a spread over a long bond, and on 18 August the 30-year Treasury yield topped 5.33%, a 19-year high. CoreWeave fell 12.7% that day.

Against that, pricing is firming. The company raised list prices roughly 25% across product lines in July, says second-quarter contracts carry contribution margins five to ten points above recent ones, and extended a 2020-vintage A100 fleet contract to 2029 — evidence the chips outlive their depreciation schedule. Customer concentration is easing but extreme: the top two customers were 65% of revenue, where Microsoft alone was about 67% in 2025. Caterpillar, Leidos and the trading firms IMC and Flow Traders were added as customers.

IREN is the counter-case, and it is not self-funded

IREN owns its power, buildings and hardware across Australia, Canada and Texas, and was a bitcoin miner called Iris Energy until late 2024. Its most recent reported quarter, to March, had revenue of $144.8m — flat year over year, and the third straight quarterly decline from $240.3m — with an operating loss of $93.3m as mining rigs came out to make room for GPUs. AI cloud was $33.6m of that. The company signed $2.8bn of new AI contracts on 20 July and lifted its exit-2026 annualised run-rate target above $4bn; in mid-August Microsoft accepted the first 50MW building at Childress, Texas, one of four under a five-year deal worth about $9.7bn.

So the shares fell 26.3% over three months while the price paid per dollar of trailing gross profit rose, from 27.57x in early May to 36.84x, because the gross profit shrank. And the leverage-free framing does not survive contact with the filings: IREN closed a $3.0bn convertible at a 1.00% coupon in May and says it has secured $9.3bn in eight months from prepayments, converts and GPU financing. That is cheaper debt than CoreWeave's, not an absence of it. Nvidia is both customer and financier, having signed a $3.4bn contract alongside a $2.1bn milestone-linked investment.

The controls disagree with each other

Nebius, the Amsterdam AI cloud spun out of Yandex, moved almost tick-for-tick with CoreWeave through August, including a 29.8% session on 12 August — the same day CoreWeave rose 19.1% on its results. Its March-quarter gross margin collapsed to 20.9% from 69.9% as GPU depreciation landed, the identical mechanism. Hut 8 broke ranks entirely, down 26.4% over the same 30 sessions; it leases 949MW of contracted capacity funded with $7.5bn of non-recourse, fully amortising project notes, and its quarterly revenue is still only $74.9m. The bounce was not one trade.

The August advance was a trough-to-peak artefact — CoreWeave bottomed at $60.82 on 29 July, ran 75% to $106.70 on 17 August, then gave back 17.7% in four sessions. Measured end to end, it is up 6.3% in a month and IREN 1.5%.

The setup

Where it stands — CoreWeave's backlog and power both grew sharply while its multiple on gross profit halved; IREN's re-rating still rests on unfiled numbers. Would confirm — CoreWeave's third-quarter adjusted operating margin moving toward the guided low teens with interest expense inside the $860-940m range. Would invalidate — Pre-depreciation gross margin falling below 65% again, or fourth-quarter active power missing the 1.85GW target. Watch next — IREN reports full fiscal-2026 results on 27 August, its first quarter containing the Microsoft ramp. Valuation — CoreWeave: 9.37x trailing gross profit, 5.51x forward, against 16.40x in early May. IREN: 36.84x trailing, 9.91x forward.

Blackstone, the world's largest alternative asset manager, told investors in July that the total value of its data-center platform had reached $185bn, up from $130bn at the start of the year. The firm holds entitled land with secured power for roughly 15 gigawatts of capacity, enough to support about $200bn of development, and expects to lease three times more capacity this year than in any year before. Nine of its ten largest markups in the June quarter were AI-related.

This is the layer of the AI build that rarely gets named: not the chips or the turbines, but the capital that buys them. For eleven months the market treated that layer as a damaged business.

What the three actually reported

Blackstone, which raises third-party money into real estate, credit, private equity and infrastructure funds and earns fees and carry on it, grew June-quarter revenue 28.5% year on year while operating income rose 54.2%. Fee-related earnings reached $1.8bn, up 22%. Assets under management hit a record $1.35tn.

Apollo Global Management is the balance-sheet case. It owns Athene, an annuity writer, and funds lending off that insurance float. Fee-related earnings of $785m set a record, up 25%, and the net spread Athene earns over its cost of funds widened to 114 basis points from 97 the prior quarter. Origination hit $74bn in the quarter — three quarters of it investment grade, averaging BBB+ at 280 basis points over Treasuries.

Ares Management, a direct lender to small and mid-sized companies, is the awkward one. Reported revenue grew just 5.8%, down from 40.8% the prior quarter, and gross profit fell 64.8%. On the fee line it looks nothing alike: fee-related earnings rose 20% to $491m at a 42.3% margin, and the firm raised a record $36bn. Its credit book shows no strain — direct-lending non-accruals below 2%, with borrower EBITDA growing 9%.

Why the money rung matters now

The demand is structural. Coverage on hyperscaler public bond deals collapsed from roughly five times in February to under two times by July, pushing issuers toward private vehicles that banks cannot underwrite at size. In August, Apollo and Blackstone arranged a $35bn financing for Broadcom's AI XPV Platform, a vehicle that buys Google Tensor Processing Units and leases them to Anthropic, with Broadcom guaranteeing residual value on $30bn of the debt. Apollo says it was the only lender able to commit the full amount, closing in weeks. Days earlier Nvidia signed memorandums with Apollo, Blackstone, Brookfield, KKR, BlackRock and Goldman Sachs to mobilize over $500bn of third-party capital — an attempt to make GPUs a borrowable asset class.

What broke them, and what didn't

The damage was not about rates. The worst sessions of the past year fell in February, when retail redemption queues broke: Blue Owl halted withdrawals, Ares capped its Strategic Income Fund at 5% after requests hit 11.6% of assets, and Apollo honored roughly $730m of more than $1.5bn requested at its debt fund in March. Ares fell 11.2% on 5 February. Peak-to-trough, Apollo lost 41%, Blackstone 46%, Ares and KKR 48% each. Credit mix explains the ranking: Apollo holds 86% of fee-earning assets in credit, Ares 66%, Blackstone only 34%.

The rate story inverts on inspection. Through the week the 30-year Treasury touched 5.323%, a 19-year high, on 18 August, Blackstone rose 2.0% and Ares was flat, while Goldman Sachs and Morgan Stanley — the supposed beneficiaries of a steep curve — fell 3.0% and 2.9%.

The shares turned on 12–13 August and have held an uptrend for eight sessions, Blackstone's first since October and Ares's since September. Over thirty days Ares is up 18.1% and Blackstone 16.7%. Over twelve months they remain down 22.0% and 15.9%, against Goldman up 37.5% and the S&P 500 up 18.5%.

Whether the price leaves room

Forward earnings is the only lens that compares these three; Apollo consolidates Athene's insurance premiums, which distorts every revenue-based ratio. Blackstone trades at 24.0x forward against 31.9x trailing. Ares trades at 24.1x forward against 61.7x trailing, with a 1.8% free-cash-flow yield — its case depends entirely on consensus GAAP earnings roughly doubling. Apollo is the outlier at 15.1x forward, on a 10.5% free-cash-flow yield.

The telling comparison is against early May, when Blackstone traded at 30.8x trailing, Ares at 64.8x and Apollo at 22.1x. Today's figures are flat to lower while fee earnings compounded above 20%. August's rally re-rated nothing.

The setup

Where it stands — Fee earnings at all three grew above 20% while trailing multiples sit at or below their early-May levels. Would confirm — Apollo delivering its guided 20%-plus full-year fee-related earnings growth with origination above $70bn again in Q3. Would invalidate — Non-traded fund redemption queues reopening above the February peak, or Ares direct-lending non-accruals passing 3%. Watch next — Third-quarter results in late October, with Blackstone's data-center platform value and Ares's Q4 digital-infrastructure fund close. Valuation — Blackstone 24.0x forward vs 31.9x trailing; Ares 24.1x vs 61.7x; Apollo 15.1x vs 28.7x.

Linde, the world's largest industrial-gas supplier and the company that pipes ultra-high-purity nitrogen, oxygen, argon and hydrogen over the fence into semiconductor plants, told investors in late July that its backlog of contracted gas-supply projects had reached a record $8.1bn. It rose by $1bn in a single quarter. The additions were not vague: roughly $1bn of work for advanced-node fabs in the western United States and a Taiwan joint venture worth about $800m for air-separation units plus hydrogen production. Electronics was the fastest-growing end market in the business, volumes up 18% from a year earlier. Group revenue set a record at $9.29bn and management raised full-year earnings guidance to $17.70–$17.90 a share.

The shares have gone nowhere for a year.

The molecule rung

This is an unusual business model dressed as a commodity. Linde builds an air-separation plant on the customer's site and signs a supply contract that typically runs fifteen years on take-or-pay terms — volume is contracted before the fab is energized, and the plant earns something close to a regulated utility's return without the regulator. Projects must clear a double-digit unlevered return hurdle and ramp over two to three years. That is why a backlog number matters more here than an order book does at a machinery company: it is already-sold revenue with a start date.

Advanced logic and high-bandwidth memory consume more gas per wafer than the nodes they replace, and every new fab and advanced-packaging line needs its own plant. So the gas majors should be a clean derivative of the same buildout that carried the equipment makers they sit beside — Lam Research, Applied Materials and Teradyne each more than tripled over the past twelve months, with Ultra Clean Holdings up 239.5%.

None of that reached the gas layer. Linde is up 1.4% over twelve months, Air Products 3.9%, and Air Liquide — the French major, the closest thing to a control on the same customers — is down 8.6%, the worst of the three, despite a record €6bn backlog and more than €1bn of electronics investment decisions in the first half alone, about 1.5 times its total for all of 2025, including a €200m carrier-gas contract with SK hynix for high-bandwidth-memory packaging.

The price of a dollar of gross profit

Earnings multiples are useless for this comparison. Air Products' trailing price-to-earnings ratio reads -1,467x, an artefact of a loss-making window. Gross margins differ too widely — roughly 47% at Linde against 33% at Air Products — for revenue multiples to compare anything. What travels is the price paid per dollar of trailing gross profit.

Linde's has compressed without interruption: 17.1x a year ago, 15.8x six months ago, 15.0x three months ago, 13.9x now. Its own fiscal year-ends read 17.5x for 2023 and 16.7x for 2024. The denominator is not shrinking — trailing gross profit went from $13.32bn to $16.18bn. Forward earnings sit at 27.3x against 31.3x trailing, at or just under the roughly 28x premium the shares have historically commanded.

Air Products has not moved at all. Its multiple was 16.8x at fiscal 2023 year-end, 16.9x in 2024, 16.1x in 2025 and 16.82x today. A flat multiple on a flat base: gross profit grew 5.1%, to $4.04bn. That leaves the slower compounder carrying a 21% premium to the faster one. Linde also retired 1.9% of its diluted shares over the year; Air Products' count is unchanged at 222.8m, and management said buybacks only come into view in late fiscal 2027 once capital spending falls.

Why each one is being ignored

The two cases are not the same. Air Products' flat year is self-inflicted and largely accounted for: it cancelled the Louisiana Clean Energy Complex and the Casa Grande hydrogen facility and booked a pre-tax charge of up to $2.9bn, producing a $2.1bn quarterly operating loss. Beneath it, the gas business is intact and improving — adjusted earnings of $3.47 a share, up 12%, an operating margin of 25.6% up more than 100 basis points, and guidance raised to $13.39–$13.49. Capital spending is being cut to about $3.5bn this fiscal year and roughly $3.0bn next. Its own backlog is near $3bn, two-thirds of the pipeline skewed to semiconductors, anchored by named sockets: on-site nitrogen, oxygen, argon and hydrogen plants for Samsung's new Pyeongtaek fab, phasing in from 2028, and supply to TSMC's Arizona 2nm line.

Linde's problem is smaller and closer to home. The margin shortfall that knocked the stock in late July came almost entirely from Lincare, its US home-oxygen unit, a roughly $45m quarterly drag from labor inflation and reimbursement policy — about $180m annualized against a group earning a 29.5% adjusted operating margin. Management is weighing a sale. Excluding it, Americas margins would have risen.

Helium separates them further. Strikes on Qatar's Ras Laffan and the Hormuz disruption removed an estimated 30% or more of global helium production, taking spot prices from about $300 per thousand cubic feet to $600–900. Linde captured strong pricing but lost the benefit to shipment dislocation costs. Air Products drew 40% of its quarterly helium volume from its own Texas cavern and took a 2% headwind where it had guided 3%.

The common factor is not demand

One mechanism fits all three: these are the most bond-like equities in the chip supply chain — long-lived assets, contracted cash flows, dividends. On 18 August the 30-year Treasury yield topped 5.33%, a 19-year high. That same session Linde's 50-day average crossed below its 200-day, ending an uptrend it had held since spring. Air Products, oddly, moved the other way after its guidance raise and its 50-day sits above its 200-day. The demand evidence points one direction and the discount rate points the other.

The setup

Where it stands — Linde's gross profit is compounding at 21% while its multiple sits at a three-year low; Air Products' multiple has not moved in three years.

Would confirm — Linde ends 2026 with backlog still above $8bn after roughly $1.3bn of project start-ups, and electronics volumes hold double-digit growth.

Would invalidate — Third-quarter earnings below the $4.45 guidance floor, or backlog falling under $7bn as start-ups outpace new wins.

Watch next — Linde's third-quarter results in late October; Air Products' fiscal fourth quarter and fiscal 2027 capex guide in November.

Valuation — Linde 13.9x trailing gross profit, 13.6x forward, against 17.1x a year ago; Air Products 16.8x trailing, 16.6x forward.

Five9 sells the software that runs corporate call centers — routing voice, chat and email to human agents — and it bills for it by the seat. In the June quarter it sold considerably more artificial intelligence. Revenue from AI products rose 78% to roughly $39m, an annual run rate above $150m, and now accounts for about 15% of subscription revenue against 9% a year earlier, according to the company's second-quarter disclosures. Management raised its full-year AI growth target to at least 60% and lifted revenue guidance to $1.266bn-$1.272bn.

It also earned less on each dollar. Gross margin fell to 53.4% from 54.9% a year earlier and 55.9% in the prior quarter, so gross profit grew 7.3% while revenue grew 10.3%. The adjusted figure tells the same story, slipping to 61.4% from 63.0%. Inference is a cost of goods sold. Selling an AI agent is not the same business as licensing a seat, and the reported margin now says so.

The seats are not disappearing

The widely held fear is that each automated conversation deletes a billable human. That is not yet visible. Five9's last-twelve-month subscription dollar-based retention was 107%, and management said concurrent agent seat counts are growing in line with contact-center revenue. Total revenue growth has accelerated three quarters running, to 10.3%. A Fortune 100 financial-services win worth about $100m in contract value will ramp to $25m of annual recurring revenue.

RingCentral, which sells cloud business telephony and is ranked first in unified-communications cloud PBX seats by Synergy Research, is monetizing the same idea from the other end. Customers using at least one paid native AI product now represent about 13% of its $2.8bn of annual recurring revenue, double a year ago, and its AI Receptionist product passed 16,000 paying customers, up 400%. Unlike Five9, its gross profit grew faster than revenue — 7.0% against 5.9% — at a 71.9% margin. Operating income rose 36% to $50.3m and net income tripled to $39.1m. It also raised the dividend 67% and cut leverage to 1.5x net debt to EBITDA, with gross debt heading toward $1.00bn from $1.64bn at the end of 2022.

Zoom is the control, because it has disclosed nothing. Its last report, on 21 May, showed revenue of $1.239bn up 5.5%, a 77.9% gross margin and operating income up 28.5%; enterprise revenue grew 7.2% while online monthly churn worsened to 3.0% from 2.8% and paid AI Companion users grew 184%. Its next print lands on 25 August.

What the price is now paying

Because gross margins across these businesses differ by more than 20 points, the comparable lens is what a buyer pays for a dollar of trailing gross profit. Five9 costs 3.81x, against 2.58x in May and 2.36x in February — a 61% re-rating over six months in which trailing gross profit grew 4.9%, from $628.0m to $658.6m. RingCentral has gone from 1.90x to 3.11x, up 64% on 3.6% gross-profit growth, though it remains the cheapest name here at 13.3x forward earnings with an 11.6% free-cash-flow yield. Zoom has moved only from 7.16x to 8.25x, and carries the lowest enterprise value to EBITDA of the group at 11.0x — but its forward price/earnings of 17.8x sits above its 15.5x trailing multiple, meaning reported earnings are flattered by non-operating gains and consensus expects earnings per share to fall. Consensus sees no acceleration anywhere: Five9 near 10%, RingCentral 5.1% then 4.6%, Zoom 4.8% then 4.0%.

The month's gains were not earned the same way. Strip Five9's two best sessions — the 19.8% jump the day after its 6 August print and an 18.2% late-July gap — and the remaining sessions compound to 0.8%. RingCentral keeps 28% after removing its earnings day and the session that followed; Zoom keeps 11.2% while disclosing nothing. Macro helped: July producer prices came in flat on 13 August and the S&P 500 hit a record, and on 18 August money rotated from semiconductors into software after Anthropic's $65bn run-rate disclosure landed below expectations. But NICE, the largest contact-center vendor at 22.2% revenue share, is down 27% over twelve months in which these three rose between 28% and 130%. This is not a sector bounce.

It is also not a settled market. Five9 is third at 13.1% share behind NICE and Genesys at 19.7%; NICE bought conversational-AI vendor Cognigy for $955m, Salesforce and ServiceNow put $1.5bn into Genesys, and Salesforce's own Agentforce Contact Center went generally available in February with native access to customer data Five9 must integrate to reach.

The setup

Where it stands — AI is selling into these customer bases faster than it is deleting seats, but at Five9 it is arriving below corporate gross margin.

Would confirm — Five9 gross margin stabilizing at or above 53.4% next quarter while AI revenue growth holds near 60%.

Would invalidate — Dollar-based retention falling below 100% at Five9, or paid-AI ARR share stalling near 13% at RingCentral.

Watch next — Zoom reports its July quarter on 25 August, its first disclosure since 21 May.

Valuation — Five9 at 3.81x trailing and 3.65x forward gross profit against 2.36x in February; RingCentral 3.11x against 1.90x.

Cohu builds the machines that grip a finished chip, drive it to the temperature it will face in the field, and press it onto a bed of contact pins so a tester can decide whether the part ships. For most of the last decade the parts in question went into cars and phones, and Cohu's fortunes moved with automotive inventories. That is no longer where its order book comes from.

In the June quarter, orders from computing customers rose 150% from a year earlier and accounted for 46% of all system orders. Automotive orders fell 24% over the same span, with utilization at auto and mobile test floors still stuck in the high 70s. Revenue reached $149.0m, up 38.4%, and gross margin came in at 45.4% against 34.4% a year earlier. Management raised full-year guidance to about 35% growth, from 25%, and lifted its high-performance computing revenue target to $100-110m.

What it is actually selling

The product doing the work is Eclipse, a thermal handler built for parts that dissipate serious power — graphics processors and server CPUs whose junction temperature must be held steady while they are tested at full load. Cohu names Han Precision as the incumbent it is taking share from, and argues that one Eclipse platform reconfigures across processor types and device generations, which lowers what a customer risks per capital order. A single customer placed a $26m Eclipse 6 order early in the third quarter, shipping mostly in the fourth. The qualification pipeline behind it runs to roughly $850m of annual customer spending, of which $190m is qualified across four accounts.

The packaging-specific content is smaller and newer: a high-bandwidth-memory inspection system qualified at a Taiwanese assembly-and-test house with an infrared sensor that finds cracks inside the stack, and a first $500,000 booking for optical-engine test.

Why the peers went the other way

Cohu is up 23% over three months. FormFactor, which makes probe cards — the consumable contact arrays that test chips while still on the wafer — is down 2.8%, and MKS, the supplier of vacuum, laser and electroplating-chemistry equipment behind both the fab and the substrate, is down 5.6%. Axcelis and Veeko, both weighted to mature nodes, fell around 16%.

The business news went the opposite way. FormFactor's gross margin has risen for four straight quarters to 50.7%, from 37.2%, on record DRAM probe-card revenue where high-bandwidth memory is roughly two-thirds of the total. MKS's electronics and packaging arm — the Atotech plating chemistry and laser drilling it bought with debt — grew 44% to $381m, faster than its semiconductor arm, while net leverage fell to 3.0x from 4.0x. Applied Materials, reporting record quarterly revenue of $9.1bn, told investors advanced packaging revenue will grow more than 70% in calendar 2026. The demand is not in dispute.

What moved the shares was rates and geopolitics. The 30-year Treasury yield topped 5.33% on 18 August, a 19-year high, repricing order books that sit in 2027 and 2028; days earlier, a report that a Shanghai state-backed firm had begun mass-producing domestic immersion lithography tools knocked ASML, Applied Materials, Lam Research and KLA about 7%. On 19 August FormFactor fell 6% with no company news of its own, alongside Aehr and Teradyne.

Where the price sits

Cohu's three-month gain is one day: the session after its 30 July results, when the shares rose 18.1%. Remove that and one other, and the last 30 days is deeply negative. It is also the only one of the three whose valuation went up. Cohu now costs 12.5x trailing gross profit, against 9.5x in February. FormFactor costs 21.6x, down from 28.7x in May while its trailing gross profit grew 16.5%. MKS costs 9.83x, down from 11.91x.

The difference is earnings. Cohu's trailing profit is negative, so the forward multiple of 57x on 2026 estimates carries the whole case. And the same memory scarcity lifting FormFactor's probe cards is a cost line for Cohu, which is pre-buying components at elevated prices with pass-through talks only beginning.

The setup

Where it stands — Cohu's order mix has flipped to computing, but the re-rating has run ahead of any GAAP profit. Would confirm — Third-quarter revenue near the guided $170m with gross margin around 45% and positive GAAP net income. Would invalidate — Computing orders falling back below 30% of system orders, or gross margin slipping under 43% on memory costs. Watch next — Third-quarter results, due late October, plus conversion of the $250m in active qualification. Valuation — 12.5x trailing gross profit versus 9.5x in February; forward price/earnings 57x, trailing not meaningful.

Applied Digital finished its fiscal year in May with signed leases covering 1,410 megawatts of critical information-technology load. The company, which builds and operates AI computing campuses in North Dakota and leases the power and buildings to cloud tenants, says those contracts represent about $36.2bn of take-or-pay revenue over initial 15-year terms. What it actually turned on during the quarter was one phase of one building — 75 megawatts declared ready for service. The distance between those two numbers is the entire investment case, and closing it costs money the company does not yet earn.

The gap between contracted and energized

The operating results were extraordinary. Fiscal 2026 revenue rose 183.7% to $611.3m; gross profit went from $22.7m to $157.7m. Yet the fourth quarter carried a $117.0m operating loss, and the full year lost $244.0m at the net line. Revenue beat consensus by roughly $164m and the shares fell 12.8% anyway, because total debt had gone to about $5bn from $2.7bn a quarter earlier and management put remaining build cost for the 1.4-gigawatt program above $10bn. In March a subsidiary priced $2.15bn of senior secured notes at 6.750%, issued at 98, to fund 200 megawatts. Diluted shares grew about 42% year over year.

So the share price fell 43.3% in three months, and the stock did not get cheaper. Applied Digital trades at 60.74x trailing gross profit, against 63.29x in early May: gross profit tripled roughly as fast as the price fell. That multiple sits on a -38.7% operating margin and a trailing free-cash-flow yield of -35.4%, and consensus does not model positive average earnings per share before fiscal 2029. Concentration is real too — CoreWeave holds all three leases at the first campus, 400 megawatts and roughly $11bn of committed revenue, backed since the quarter by springing parent guarantees and a $50m letter of credit after CoreWeave earned an A3 rating.

Why they all fell on the same two days

This is a business model funded by borrowing long against long leases, which makes the far end of the yield curve its cost of goods. Two sessions did most of the damage. On 29 July the Federal Reserve held rates and the Dow fell 1,153 points as the 30-year Treasury yield reached its highest since 2007. On 18 August the 30-year topped 5.33%, a 19-year high, and every one of the six US-listed data-center operators fell — VNET 16.9%, Keel 16.2%, WhiteFiber 10.8%, SharonAI 10.6%, Applied Digital 8.6%, GDS 4.9%. Across the month five of six fell, from Keel's 27.1% to Applied Digital's 9.5%.

Keel Infrastructure is the redomiciled successor to bitcoin miner Bitfarms, converting a 2.2-gigawatt power pipeline to AI. WhiteFiber rents GPUs and colocation and grew revenue 54% last quarter. SharonAI has traded only since February. All three are pipeline stories with no earnings to defend them.

The one whose price fell furthest below its numbers

VNET, a Chinese operator running wholesale campuses for internet and cloud tenants, is the divergence. Wholesale revenue grew 29.3% last quarter; capacity in service crossed a gigawatt for the first time at 1,007 megawatts, up 49.4%, with 73.9% utilized and mature sites at 92.5%. Adjusted EBITDA rose 25.4% to about $128m, with margin widening to 33.0%. The shares are down 32.3% in three months and sit 52% below their 12-month high, at 9.72x trailing EV/EBITDA. The check on that: free cash flow is deeply negative, and 2026 capex of RMB10-12bn — roughly $1.4-1.7bn — must be raised, which is exactly what a 5.33% long bond repriced.

GDS, the Shanghai-based colocation landlord, was the only member that held, and not because its quarter was good. Revenue grew 6.2%, down from 23.6% the prior quarter; gross profit fell 12.9%; adjusted EBITDA rose 2.5% on utility costs. What it has is demand and a different lender: backlog went from 450 to 757 megawatts on record first-half bookings of 470 megawatts, funded 60/40 debt to equity in an onshore Chinese bank market management called highly supportive, at 4.7x net leverage. It trades at 14.03x trailing EV/EBITDA and 1.37x book.

Applied Digital's uptrend broke in stages — its 50-day average crossed below its 200-day in late July, and by 20 August the decline was firmly established. VNET has been in a downtrend since 25 June. The businesses did not turn; the discount rate did.

The setup

Where it stands — Growth is intact at all three landlords; the long bond, not demand, repriced the equity over the past month. Would confirm — Applied Digital energizes materially more than 75 megawatts next quarter without new equity issuance. Would invalidate — A CoreWeave lease amendment, or GDS bookings falling short of its 1-gigawatt 2026 target. Watch next — Applied Digital's fiscal first quarter, due October, and its quarterly capex run-rate against roughly $600m guided. Valuation — Applied Digital 60.74x trailing gross profit, 42.67x forward, versus 63.29x in May; VNET 9.72x trailing EV/EBITDA, GDS 14.03x.

A quarter's margin, spent on GPUs

Akamai Technologies — which secures, routes and delivers websites, applications and programming interfaces from servers embedded inside internet providers' networks in more than 130 countries — spent the June quarter turning AI contracts into hardware. Capital expenditure reached $347m, against 21% of revenue a year earlier and 32% now. Management guided full-year spending to roughly 40% of revenue, with the September quarter as high as 43-46%.

That build shows up immediately in the cost line and only later in the revenue line. June-quarter revenue rose 5.4% to $1.100bn while gross profit slipped 0.5% to $613.8m, gross margin falling 3.3 points to 55.8%. GAAP operating income nearly halved to $80.3m. The shares fell 6.4% on 7 August as next-quarter guidance came in below consensus and the full-year adjusted earnings forecast was cut.

The underlying mix is not the problem. Security revenue reached $604m in the quarter, up 10%, and Cloud Infrastructure Services — the compute arm that hosts AI inference — grew 39% to $99m. Delivery, the original content-distribution business, fell 6% to $396m and is deflating at roughly 5-7% a year as hyperscalers commoditize it. Security plus compute is now 64% of revenue, a clear majority. Akamai's edge is 4,400-plus points of presence placing workloads within a single network hop of most internet users, architecture no AI startup can rent into existence.

The round trip

Price per dollar of trailing gross profit is the cleanest lens here, because reported earnings are distorted by the depreciation of assets bought ahead of the revenue they will carry. On that measure Akamai has gone nowhere and back: 6.54x on 20 February, 9.26x on 18 May after the Anthropic contract, and 6.58x today. Trailing gross profit was about $2.44bn at all three dates. Every move was multiple.

The shares gapped 27.8% higher on 8 May, peaked at $161.14 on 3 June, and now trade at $110.42 — below the $115.62 close that preceded the announcement. Management says its GPU capacity is sold out, that typical large contracts carry 60-70% cash gross margins, and that 2027 revenue should grow in the low teens against 6-8% this year. None of that is in the price. The buyback was suspended after $616m of repurchases, $3.5bn of zero-coupon convertible notes were issued, and the diluted share count still rose 5.8% to 153.7m. Two customers — Anthropic and an unnamed robotics developer on a $600m-plus deal — now anchor years of economics.

The counter-cases

F5, which sells BIG-IP application-delivery appliances and software to enterprises and governments, did the opposite. June-quarter revenue rose 10.9% to $865m and gross profit grew faster still, up 12.6%, with margin at 82.2%. Guidance was raised to 9-10% growth and the shares gained 4.0% that session. Notably, this is the company that a year ago warned of damage from a breach in which nation-state actors held access to its product-development environment for at least twelve months and stole source code; it guided to 0-4% growth at disclosure. The feared churn never printed. The caution is composition: systems revenue grew 32% while software grew 7% and subscriptions 9%, and management expects mid-to-high single-digit systems growth from here. F5 now costs 7.98x trailing gross profit, up 26% in six months against 6.3% gross-profit growth.

A10 Networks, the smallest name, is the control that spoils the tidy story. Revenue accelerated for a third straight quarter, up 15.5% to $80.1m, and full-year guidance was raised. The shares fell 9.2% anyway and are down about a quarter in a month. The objection is leverage, not demand: operating income fell 12.5% as margin dropped to 11.3%, and A10 deepened its dependence on a single hyperscaler, Microsoft.

Cloudflare, the member the market has already re-rated as an AI serving layer, grew revenue 36% to $696.1m with net retention at 120%. It is the expensive one; the incumbents are the cheap ones.

All four fell together on 17-20 August, when the 30-year Treasury yield topped 5.33%, a 19-year high — Akamai worst at -10.4%. Trend has broken in the incumbents: F5 slipped out of its strongest uptrend on 28 July, A10 on 7 August after 122 sessions. The common thread is not delivery being repriced. It is that this quarter the market paid for gross profit already earned, and Akamai converted the most backlog into capital spending and the least into margin.

The setup

Where it stands — Akamai's AI compute backlog is worth roughly nothing in the share price, because gross profit has not moved in a year. Would confirm — September-quarter Cloud Infrastructure Services revenue above $120m with gross margin stabilizing near 56%. Would invalidate — Gross margin below 54% in the September quarter with capex still above 40% of revenue. Watch next — Third-quarter results in early November, and the first 2027 revenue guide against the "low teens" framing. Valuation — 6.58x trailing gross profit and 16.4x forward earnings, against 9.26x gross profit in May and 38.9x trailing.

Accenture, the Dublin-headquartered professional services firm that sells strategy work, systems integration and outsourced operations to large enterprises and governments, spent three years telling investors exactly how much generative-AI work it had sold. Then it stopped. The last disclosure put advanced-AI bookings at $2.2bn for the quarter, up 76% year on year, on a cumulative record of roughly $11.5bn booked across 11,000 projects. Management's explanation was that the category had become so pervasive it was no longer meaningful to break out.

What arrived in its place, on 18 June, argued the other way. Revenue of $18.72bn grew 5.6% in dollars and only 3% in local currency, down from 8.3% growth the prior quarter. New bookings of $19.3bn fell 2%, leaving a book-to-bill of exactly 1.0 — one dollar of future work sold for every dollar delivered. The company cut full-year local-currency growth guidance to 3–4%. The shares fell 18% that day.

Since 22 July they have risen 32.3%, to $185.28, with no company disclosure in the interval. The stated drivers are analyst upgrades citing cheap multiples and short covering, and the average price target set over the past month, $189.88, sits barely above where the stock closed.

What the rally is paying for

Accenture trades at 14.7x trailing and 13.4x forward earnings, against 21x to 37x at each of its last five fiscal year-ends. Measured against gross profit — the fitting lens for a business whose cost of revenue is people — the price paid per dollar earned is 4.85x, versus 5.97x six months ago and 7.29x a year ago. It bottomed at 3.69x on 22 July. The re-rating has recovered barely a third of a year's de-rating, which is the strongest thing that can be said for it.

The next print will be the test. Fourth-quarter guidance of $17.75bn–$18.4bn compares with $17.60bn a year earlier: growth of 0.9% to 4.6%, slower again. Meanwhile the firm plans $9bn of acquisitions, including the operational-technology security firms Dragos, RunZero and NetRise — capability bought rather than grown. And in September 2025 it cut 11,000 jobs in an $865m restructuring, with chief executive Julie Sweet saying it was exiting people for whom reskilling "is not a viable path".

Cognizant is repricing the hour

Cognizant, the Teaneck, New Jersey firm that delivers consulting and outsourcing to banks, insurers, healthcare payers and manufacturers largely from offshore centers, is the one name here that raised anything. It lifted full-year adjusted earnings guidance to $5.70–$5.82 a share and held adjusted operating margin guidance at 16.0–16.2%. Second-quarter operating income grew 7.0% on revenue up 4.5% — real operating leverage, a sixth consecutive quarter of margin expansion.

The qualifications matter. Headcount of 356,700 is up 12,900 year on year, so the growth was delivered on about 3.7% more people, not on price. Roughly 150 to 200 basis points of it is inorganic, following $1.3bn of deals including Astreya. And gross margin slipped to 33.37% from 33.67% — where a productivity give-back to clients would first appear.

What distinguishes the company is that it is trying to sever revenue from the billed hour rather than defend it. Its TriZetto healthcare platform now clears $1.1bn a year at higher margins than the firm as a whole. Fixed-price and transaction-based work has grown for three straight years, and some clients now ask for rate cards with inference costs embedded. Management is standing up 5,000 engineers to audit agentic workflows and 10,000 operators to run blended human-machine teams. The stock's 19% trailing-to-forward earnings compression, to 10.8x, is roughly twice Accenture's.

Infosys says the word out loud

Infosys, the Bengaluru outsourcer with 328,062 employees, cut its full-year constant-currency guidance to 1.5–3.0%, which nets to something near 0.5% once acquisitions and mix are stripped out. Management confirmed genuine price deflation on large renewals, increasingly demanded mid-contract, and declined to size it. It is nonetheless hiring 20,000 graduates. Its shares rose 9.2% over the month, the weakest of the group, and the Indian leg of the move began with a Jefferies upgrade of the sector to neutral — a positioning call, not a demand one. At 7.72x trailing gross profit it is the most expensive of the three, and the only one whose forward multiple sits above its trailing.

EPAM, the US-listed engineering-services firm, cut full-year guidance to 3.2–4.2% growth and fell more than 20% on the print, yet is up 27.5% over the month regardless.

All of them now compete with their suppliers. OpenAI has launched a deployment arm with $4bn from 19 firms and bought the consultancy Tomoro; Anthropic has a $1.5bn services venture backed by Blackstone and Goldman Sachs, both selling embedded engineers instead of analysts. Cognizant claims more than 10,000 Claude-certified architects — it is Anthropic's partner and its rival at once.

Strip each name's two best sessions and Accenture is still up 16.8% and Cognizant 10.5% over the month, so this is no longer the two-session rotation it looked like in early August. In the four sessions to 21 August the two added about 8% each while Nvidia fell 4.6%. All three flipped out of a steep downtrend on the same day, 7 August — one signal, not three company turns. Over twelve months Accenture remains down 27.1% and Infosys 30.0%.

The setup

Where it stands — Accenture's shares have recovered a third of a year's de-rating on upgrades and rotation, not on any new demand disclosure. Would confirm — Fourth-quarter bookings returning to year-on-year growth with book-to-bill back above 1.0. Would invalidate — Fourth-quarter revenue near the low end of $17.75bn, or fiscal 2027 local-currency guidance below 3%. Watch next — Accenture's fiscal fourth quarter ends 31 August and reports in late September. Valuation — Accenture at 14.7x trailing and 13.4x forward earnings, against 21x–37x at its last five fiscal year-ends.

A developer that hit its marks

NexGen Energy, a Vancouver company whose only asset is the undeveloped Rook I uranium deposit in Saskatchewan's Athabasca Basin, has now cleared the hurdle that decides whether such projects exist at all. The Canadian Nuclear Safety Commission approved its environmental assessment and issued a licence to prepare the site and construct on 5 March 2026, 14 business days after hearings closed. NexGen took a final investment decision, raised roughly C$953m — about US$690m, and described as the largest equity financing in the sector's history — and awarded the shaft-sinking and underground engineering package, worth more than half the C$2.2bn (about US$1.6bn) capital estimate, at the figure it published in August 2024. Construction inflation has run hard across mining since; the number did not move. Site workforce is above 300, and first production is targeted around 2030.

The shares are down 12.3% since February. Market capitalization stands at $7.19bn against $8.01bn in mid-May, even though diluted shares rose from 570.0m to 705.3m over the year. Investors own nearly a quarter more of a materially de-risked asset for less money in aggregate. The financing is not finished: C$970m of liquidity funds well under half the build, with heavy spending deferred to early 2027 and project finance, government funding and customer prepayments all still under evaluation. NexGen has also chosen commodity exposure over certainty — 11.3m pounds committed at spot-indexed delivery prices, with 96% of the reserve base uncommitted.

Why that choice matters

Cameco, the world's largest publicly traded uranium producer and a 49% owner of reactor builder Westinghouse, is the counter-example. Its second-quarter realised uranium price rose 15% to C$93.13/lb — roughly US$68, well below a spot price near $88 and a long-term contract price stuck at $94, the highest since 2008. A fixed contract book signed in cheaper years is currently a drag on realisation, not a premium. Second-quarter revenue fell 7.2% to $814.1m, gross margin compressed to 21.1% from 29.3%, and net income fell 92% to $25.2m, with production of 3.9m pounds down 15% after spring road disruptions at Key Lake and McArthur River and a two-week suspension at Cigar Lake. Full-year production guidance of 19.5m to 21.5m pounds is unchanged.

Cameco's shares are down 15.5% over six months, yet the stock is more expensive than it was: 172.7 times trailing earnings against roughly 105x in early May, and 64.2 times trailing gross profit against 38x to 40x. Gross profit fell faster than the price. The forward multiple of 66.4x rests on 2027, not 2026, when consensus has revenue rising only 4.5%. The optionality is real — Westinghouse has confidentially filed a draft registration statement for a possible listing — but nothing has been given back on price.

The enricher's backlog outruns its shipments

Centrus Energy, the only American-owned commercial enrichment company, is where the scarcity story should be loudest. US utilities bought about 3.28m separative work units of Russian enrichment in 2025, close to 26% of their purchases, and the waiver regime ends in a full ban in 2028, with enrichment now near $160 per unit against about $40 before 2022. Centrus's backlog nearly doubled to $4.5bn through 2040, $2.4bn of it under definitive agreements, helped by a firm off-take with reactor developer X-energy. But second-quarter revenue growth of 14% came from reselling natural uranium; enrichment volumes fell 23% and gross margin dropped to 28.3%. New Western capacity is coming — Urenco is expanding its New Mexico plant by almost half — but not before 2029. Centrus's forward price/earnings of 74.6x sits above its trailing 74.2x, the reverse of a growth discount, because 2026 consensus earnings of $2.50 are 44% below last year's $3.90.

The month's trading points the same way. Over 30 days the miners led — Energy Fuels up 23.0%, Uranium Energy 22.5%, NexGen 15.0%, Cameco 13.4% — while Centrus added 6.9% and the reactor side fell, BWX Technologies down 10.5%. On 21 August the group gained 6% to 9% in a session with no company news and the S&P 500 up 0.2%. Cameco's 50-day average nonetheless sits below its 200-day for the first time in over a year. Money is moving down the stack, toward pounds — which is precisely the exposure NexGen has chosen and has not yet financed.

The setup

Where it stands — NexGen has permits, a budget held at estimate and a funding gap; Cameco and Centrus have de-rated without getting cheaper on profits. Would confirm — NexGen closes project financing covering the balance of the C$2.2bn build without a further equity issue. Would invalidate — A revised Rook I capital estimate above C$2.2bn, or first production guided past 2030. Watch next — Third-quarter results in November; freeze-system commissioning due early 2027 before heavy construction spend begins. Valuation — NexGen 5.38x book, no earnings multiple; Cameco 172.7x trailing and 66.4x forward against ~105x in May.