Notes on the market.

The AI buildout moved another notch from announced intent to signed, underwritable paper this week, and the counterparties writing the checks are the kind every reader on this desk can underwrite: lenders, bond buyers, and an AI lab committing to eleven-plus billion dollars of compute capacity. The clearest single signal is a cluster of 8-K filings in which data center owners closed debt or extended contracted revenue, each with a dollar figure attached.

On the borrower side, Hut 8 Corp. announced the closing of a $1.07 billion four-year senior secured revolving credit facility, which the company stated strengthens its parent-level liquidity and expands funding capacity across the development lifecycle sec.gov. CleanSpark, Inc. announced that its wholly owned subsidiary CSDC Finance I, LLC closed its offering of $2.276 billion aggregate principal amount of 7.875% senior secured notes due 2031 sec.gov. Cipher Digital Inc. announced a series of agreements that extend the contracted lease term at its Barber Lake data center in Colorado City, Texas from 10 years to 20 years and increase total contracted revenue at the facility from $3.8 billion to $9 billion across the additional term, adding roughly $5.2 billion of contracted revenue sec.gov. For a lender, these three are the same story told three ways: revolver, notes, and contracted lease revenue all moved on the same day.

The compute buyer side is signing at matching scale. Data Center Dynamics reports that Akamai signed an $11.6 billion compute deal with Anthropic, under which the AI firm leases capacity from the cloud and edge provider to support CPU workloads datacenterdynamics.com. The same outlet reports that UK data center firm Virtus secured £2.45 billion in financing to fuel projects in the UK and mainland Europe datacenterdynamics.com. Atlas Energy Solutions Inc. disclosed in an 8-K that it executed separate cost reimbursement agreements with a leading frontier AI lab alongside equipment purchase agreements to secure long-lead-time equipment and incremental power generation equipment for specific data center projects, with the agreements allocating that equipment to a specific customer and facilitating near-term financing sec.gov. MARA Holdings, Inc. disclosed in an 8-K that a subsidiary acquired the membership interests of a project company from HIF USA LLC, retaining an equity interest for the seller sec.gov. Volato Group, Inc. reported in an 8-K that a large AI customer placed orders representing $1.17 billion in aggregate contract value for AMD GPU infrastructure at Alignment Engine's campus in Ohio, with deployments running on AMD MI355X and MI455X GPUs and a contractual start date of December 31, 2026 sec.gov.

The regulator is writing the rulebook for the power those facilities draw. In the Federal Register, the U.S. Nuclear Regulatory Commission, as lead agency, and the Department of Energy Office of Energy Dominance Financing, as a cooperating agency, are issuing an environmental assessment and finding of no significant impact evaluating the effects of proposed federal actions related to reauthorizing power operations at the Christopher M. Crane Clean Energy Center federalregister.gov. For anyone pricing a facility's time-to-energize, this document is where the schedule gets set.

A second thread ran through the wire this cycle, on the software and automation layer that the buildout supports. HCLSoftware announced it will acquire Robotiq.ai, a move it said adds enterprise RPA capabilities to its HCL UnO Agentic platform and strengthens end-to-end orchestration across AI agents and enterprise applications prnewswire.com. Duck Creek launched an agentic FNOL solution that uses AI agents to capture, validate, and route insurance claims across channels prnewswire.com. Virtana was named a winner of SiliconANGLE's 2026 TechForward Awards in the AI-Driven Cloud Infrastructure category businesswire.com. The Information reports that startups Fal and Fireworks AI, which sell access to AI models and servers, are weighing new funding rounds as inference demand soars, with Fal discussed at a $15 billion valuation theinformation.com.

The watch item is whether the freight of new debt, contracted lease revenue, and committed compute keeps clearing at these prices once the first facilities energize. Axios reports that AI companies, private equity firms, and unions are banding together in a multimillion-dollar coalition, the American Infrastructure Alliance, to push for responsible data center growth in seven states axios.com, while Utility Dive reports that a study by the Environmental Protection Network argues federal actions are allowing on-site backup generators at data centers to release emissions with fewer checks utilitydive.com. A buyer, a banker, or a lender watching this lane will want three numbers on every next deal: the contracted lease revenue, the senior secured coupon, and the energized-power date. All three moved in public filings this cycle.

The money in the AI economy this week is not chasing the next foundation model. It is financing the data centers, the fiber, and the power that models run on, and the transactions being signed are the kind an owner, a banker, or a lender can underwrite with a number attached. The clearest evidence is a British neocloud securing $3.36 billion in convertible financing ahead of its US IPO. TechCrunch reports that the funding comes from Third Point, Nvidia, and others, and will fuel the company's massive AI data center buildout techcrunch.com. Data Center Dynamics confirms the same figure and notes it follows publication of the company's IPO prospectus datacenterdynamics.com. That is a borrower raising capital against contracted compute, not a lab raising capital against a demo.

The 8-K filings tell the same story from the lender's side. CleanSpark, Inc. announced that its wholly owned subsidiary CSDC Finance I, LLC closed an offering of $2.276 billion aggregate principal amount of 7.875% senior secured notes due 2031, in a press release filed as an 8-K exhibit on September 25 sec.gov. Cipher Digital Inc. announced a series of agreements that extend the contracted lease duration of its Barber Lake data center in Colorado City, Texas from 10 years to 20 years and increase total contracted revenue at the facility from $3.8 billion, adding roughly $5.2 billion of contracted revenue under a $9 billion additional 10-year term sec.gov. Atlas Energy Solutions Inc. disclosed that it executed separate cost reimbursement agreements with a leading frontier AI lab alongside equipment purchase agreements to secure long-lead-time equipment and incremental power generation equipment for specific data center projects, with the agreements allocating that equipment to a specific customer and facilitating near-term financing sec.gov. Each filing states what was signed and for how much; none of the three is a narrative about capability.

The supplier and grid layer is moving just as fast. Data Center Dynamics reports that Vertiv will acquire data center fluid management firm King Environmental Services, with the transaction expected to close before the end of the year datacenterdynamics.com. The same outlet reports the U.S. Department of Energy is unveiling $1.9 billion in funding for 31 grid upgrade projects aimed at speeding data center connections, a program expected to unlock 23 gigawatts of additional capacity on the grid datacenterdynamics.com. Vocus is building a 4,000-kilometer fiber link between Brisbane and Darwin to support demand from hyperscaler, neocloud, and enterprise customers datacenterdynamics.com. Power, cooling, and routes are the new collateral.

The regulator is writing the rulebook for that collateral at the same time. In the Federal Register, the U.S. Nuclear Regulatory Commission and the Department of Energy Office of Energy Dominance Financing are issuing an environmental assessment and finding of no significant impact evaluating the reasonably foreseeable environmental effects of proposed federal actions related to reauthorizing power operations at the Christopher M. Crane Clean Energy Center federalregister.gov. The NRC separately published a proposed rule aimed at regulatory enhancements for reactor licensing, decommissioning, and operational oversight, conducted consistent with Executive Order 14300 federalregister.gov. For anyone sizing a data center loan, these are the documents that decide how fast and how expensively a site can actually energize.

Two final reads put a lens on the cost of that buildout. The Information reports that Oracle, set to lease a New Mexico data center on behalf of OpenAI, sent a notice to the project's developer citing force majeure as a way to assert contractual rights to withhold payments if project delays persist theinformation.com. A separate Futurum Research report, carried on Business Wire, finds agentic AI can raise token consumption per task by 10 to 100 times versus a simple inference call, accelerating the shift away from per-token pricing businesswire.com. The two together name the risk that sits on every one of these facilities: the power and cooling have to be financed before anyone is certain the usage that pays for them will actually arrive on schedule.

The watch item for the next cycle is whether the neoclouds' converts and the operators' lease extensions and note offerings hold up as contracted revenue, and how quickly the DOE's grid money translates into energized capacity. A lender pricing a hold on any of this wants three numbers: contracted lease revenue, the senior secured coupon, and time-to-power. All three moved in public filings this week.

On August 28, 2025, General Catalyst laid out the core shift in The Future of Services: customers care about outcomes, not products. The essay argues the SaaS gold rush made “productized” the only legible category, while services businesses — long dismissed as low-margin and VC-taboo — deliver outcomes on a much bigger scale than software products alone. For an operator who still runs the recap, the sale, or the loan, this is the through-line of today's cycle: the unit of value is shifting from licensed software toward purchased outcomes, and every downstream asset has to be re-read in that light.

Caritas Venture Co.'s 2026 market map caritas.ventures names the vehicle for that shift. It describes a new category of firm that is “buying, or economically aligning with, established businesses and transforming them with embedded AI teams,” and reports that it reviewed 80+ owner-operators, acquisition platforms, and transformation firms, grading each on public evidence. That is the AI-native private-equity rollup: mature businesses acquired and re-run with an embedded AI function rather than a bolt-on software line. The relevant screening question is no longer whether a firm sells software, but whether it can repurchase an outcome and attach an embedded AI team to the operating rhythm.

On the capital side, NVIDIA's August 11, 2026 post by Jensen Huang blogs.nvidia.com announced partnerships with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR to establish independent financing platforms designed to mobilize over $500 billion of third-party capital to support AI infrastructure buildouts. Huang's framing is precise: the industry has moved from buying chips and building data centers project by project to financing AI factories as productive infrastructure, with repeatable platforms and long-term investors. For anyone underwriting a loan or a recap against compute economics, that is a structural change in how the underlying asset is capitalized, not a one-off funding round.

The credit backdrop is maturing in parallel. Wellington Management wellington.com describes private credit entering a more mature phase marked by greater selectivity, broader opportunities, deeper market infrastructure, and growing emphasis on manager selection. Moody's moodys.com traces the same structural expansion — from non-bank lending to middle-market borrowers outward into private placements, asset-based finance, real estate, infrastructure, and fund finance — while warning that as capital shifts into private markets, opacity increases, risk profiles diverge, and traditional public-market signals become less reliable. For a lender pricing a hold or an operator running a sale, that divergence between stale public signals and private reality is exactly the risk that needs a named borrower and a named benchmark rather than a sector-level call.

Two measurement anchors close the loop. Thomson Reuters' 2026 AI in Professional Services Report thomsonreuters.com frames AI adoption inside law, tax, and audit firms as the productivity layer of the services economy — the report explicitly rests on 175 years of Thomson Reuters knowledge rather than a speculative demo. The Cliffwater Direct Lending Index cliffwaterdirectlendingindex.com is the public benchmark for the private-credit leg, and the BEA's GDP by Industry series bea.gov is the macro ground truth for where output is actually accruing. Together these give an operator three independent checks: adoption inside the service firms (Thomson Reuters), the marked-to-market private-credit leg (Cliffwater), and the sector lens on real output (BEA).

Forecast (12m/24m). Twelve months out: the 80+ AI-native PE map consolidates toward a smaller set of platforms that can show publicly-graded evidence rather than a stated thesis, and private-credit selectivity widens the spread between top managers and the general field. Twenty-four months out: AI-factory financing platforms mature into a durable, repeatable asset class that detaches infrastructure capex from any single operator's balance sheet, and professional-services AI moves from report-stage adoption toward priced outcomes, putting pressure on the unit-economics benchmarks inside law, tax, and audit firms. Recurring pattern this cycle: fetch returned 8 of 9 pinned primaries with HTTP 200; Bessemer's owning-the-outcome URL returned HTTP 404 and was excluded with nothing invented; no coverage gap.

Two curves are moving in opposite directions across the field this cycle, and the gap between them is the story. The capital machinery around AI infrastructure and AI-native services is maturing quickly — it is being financed, indexed, and re-allocated from plain corporate loans toward asset-backed and infrastructure credit. The proof that any operator running an AI-transformed services business has delivered a verified outcome is still empty. The larger the first curve gets, the more the second one matters.

The prize is the services economy, and it is barely moving. General Catalyst puts the scale in one sentence: U.S. service industries generate more than $6T annually against roughly $370B for the entire software market, and the firm's "AI-enabled roll-up" model pairs applied AI with outright acquisitions of services businesses to chase software-era Rule-of-40 economics generalcatalyst.com. The government tape agrees on scale and adds velocity. In the first quarter of 2026, real GDP rose at a 2.1% annual rate on the third estimate, but the composition was lopsided: real value-added grew 7.5% for government and 4.5% for private goods-producing industries, against only 0.8% for private services-producing industries bea.gov. The biggest sector is the slowest-growing one — exactly the untouched-by-AI opportunity the roll-up thesis claims to be filling.

The outcome is now the product, but availability is not opportunity. Bessemer marks the shift from the experience layer to the delivery layer: in the cloud era software won by becoming the system of record, but now AI can deliver the work itself, so "the outcome is the product," and the framework's stated warning is that agent availability does not equal market opportunity bvp.com.

The proof tier is still empty. Caritas reviewed 80+ owner-operators, acquisition platforms, and transformation firms, then graded each on public evidence while excluding demos, projected margins, and branded "AI OS" claims; across all four columns of its 2026 grid, the Grade A "third-party verified" row reads none yet.

The capital is maturing even as the proof lags. Moody's 2026 outlook sees private-credit assets under management exceeding $2T in 2026 and approaching $4T by 2030, with the mix shifting from corporate lending to asset-backed finance and into newer pools such as consumer loans and data-infrastructure credit moodys.com. Discipline is arriving at the same time. Wellington describes a maturing market marked by greater selectivity and manager focus, and its cited Cliffwater Direct Lending Index data shows the 2021 and 2022 origination vintages carried the largest non-accrual balances as of Q1 2026 — more than three times the 2024 vintage wellington.com. That index is the field's read on how the money is actually performing, and the answer is uneven: the loosest vintages are paying for it.

Compute is becoming collateral, but power is the constraint. NVIDIA has announced partnerships with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to build independent financing platforms designed to mobilize over $500B of third-party capital for AI infrastructure, with residual-value support of up to 25% of an opportunity on a project-by-project basis blogs.nvidia.com. Moody's separate "Power without delivery" note warns the AI boom is increasingly constrained not by compute demand but by power availability, transmission infrastructure, and energization timelines, and that the gap between a physically completed data center and operational readiness is becoming a financing risk moodys.com.

Adoption is real, but the ROI is unmeasured. Thomson Reuters' 2026 report puts organization-wide generative-AI use at 40%, up from 22% a year earlier, while only 18% of professionals say their organizations track ROI; agentic AI sits at 15% in use with another 53% planning or considering it, and 77% expect it to be central to their workflow by 2030 thomsonreuters.com. The pattern is the same one the empty Grade-A row exposes: usage is spreading faster than proof.

NET READ. The scaffolding of a real asset class is being assembled — an index, financing platforms, and a re-allocation toward asset-backed and infrastructure credit — while the operating proof that AI-native services owners deliver verified outcomes is still absent from the top of the grade curve. Over the next 12 to 24 months the binding question is not whether capital keeps flowing; it is whether a single operator occupies the empty verified tier and puts a number under the claim.

Case-against. An empty verified tier can mean "too early to verify" rather than "no value," and a 0.8% services output figure could mark a sector already near its ceiling of slow growth rather than a giant untapped pool. On that read, the $500B+ platform mobilization and the asset-backed rotation are capital moving to infrastructure with obvious collateral, not proof that AI-transformed services firms have earned a place in it. Both readings share one discipline: watch the non-accrual vintages and the Grade-A row — the two places where the story either confirms itself or quietly disappoints.

1. The Future of Services
2. Owning the outcome
3. AI Rollups & AI-Native PE: 2026 Market Map
4. Private credit outlook
5. Private Credit Outlook 2026
6. Power without delivery
7. AI Factory Compute Is Becoming an Investable Asset Class
8. 2026 AI in Professional Services Report
9. GDP by Industry
10. cliffwater.com

The services economy this cycle turns on one question: where does durable value live when the actual work behind a professional service can be done by a machine rather than a person? The field sources converge on the delivery layer. Bessemer makes the distinction clearest: technology has always touched services at the experience layer (better funnels, onboarding, interfaces) but never the delivery layer, and AI changes delivery itself bvp.com. General Catalyst says the same from the operator side, arguing that builders capture more value by directly owning and operating the service businesses that deliver outcomes, not by selling licenses for tools generalcatalyst.com. Caritas maps who is actually doing this, reviewing 80-plus owner-operators across four paths and grading each on the strength of its public evidence caritas.ventures. And Wellington frames the financing side, where private credit is shifting from broad corporate enthusiasm into a mature, selective market that diversifies across collateral types and leans on manager capability wellington.com.

12-MONTH FORECAST

Inside twelve months, expect the outcome-tiering to harden. Caritas scores every firm on whether its named outcome is verified by a credible third party (A), only company- or investor-reported (B), or a real thesis with no publicly verifiable outcome yet (C), and the A tier remains empty. Watch that gap as the checkpoint: the operators who move into third-party-verified territory will separate from the sea of B-grade claims. On the financing side, expect Wellington's dispersion to widen: continued redemptions from semi-liquid funds and pockets of borrower stress will force investors to pick managers rather than the asset class. Meanwhile the Bessemer surplus question becomes binding for founders: AI that delivers at software-like margins only matters if the surplus is not competed away and if something (recurring revenue, compounding data, a regulatory moat) stops the work from leaving.

24-MONTH FORECAST

On a 24-month horizon the delivery-layer thesis compounds into a two-speed market. Bessemer's long-horizon agents flips the cost equation, with hours-long tasks running at the variable cost of inference, and General Catalyst's roll-ups are already demonstrating the economics: Crescendo automating 80-percent-plus of interactions, Dwelly doubling EBITDA margins across six acquired agencies, Eudia attacking a 1.05-trillion-dollar legal industry that still equates time with value. The services sector is roughly 13 percent of US GDP, about ten times the size of software, and it remains largely untouched by meaningful AI integration, which is the whole size of the opportunity and the risk. Watch whether any of Caritas's mapped firms crosses to Grade A within the window, because until one does, the entire category stays early and capital-fragile, and private credit keeps maturing toward scrutiny rather than enthusiasm.