ClarityBeta

Investment & Market Intelligence

The Investor

10 min read · 2,033 words · 3 moves to watch
FRIDAY, OCTOBER 2, 2026
The Signal

CoreWeave's lenders charge +5.5% for the years no customer has signed, versus +2.25%.

The ring-fenced $8.5B project loan gets the cheap rate and the $2.6B facility gets the expensive one, a gap of about $85M a year, which is what lenders now charge for contract risk. The banks that anchored $63.1B of Oracle and Stargate debt are getting choosier just as memory lifts Nvidia system prices about 17%, so the next borrower pays more for the hardware and more for the money. This could be ordinary discipline rather than a retreat, but the pricing says otherwise.

In Play

  1. AI Infrastructure Debt Is Priced by Seat, Not Layer

    Lenders charge CoreWeave benchmark +2.25% on a ring-fenced project loan and a 9.625% coupon on unsecured notes, according to Artificial Intelligence Made Simple's review of its credit documents. The Information reports that SocGen, SMBC and MUFG, which together anchored $63.1B of Oracle and Stargate debt, are now more selective. For your book, AI infrastructure risk is priced by where a claim sits in the capital stack. Equity behind debt that outlasts its contracts is the first-loss position.

  2. Memory Makers Now Capture the AI Margin

    The Information reports that Micron's free cash flow reached $33B in the quarter ending Sept. 3, more than Apple's $32B in its July quarter. Micron's revenue nearly quintupled to $54.2B, and its gross margin hit 86.8%. Nvidia system prices are up about 17%, partly because of memory costs, and Micron expects RAM shortages to tighten through 2028. For your neocloud and GPU-cloud exposure, the cost of each new rack is rising while lenders get choosier.

  3. Continuation Vehicles Set 'Par' Without a Market Test

    Single-asset continuation vehicles grew 88% to $34B in H1 2026, about 80% of all secondaries growth, according to Augment's analysis of Evercore data. Venture LP stakes sold at 79% of NAV, while 86% of single-asset CVs priced at or above NAV. Much of that gap is a control premium, because the GP picks the asset, the buyer and the NAV. ILPA's draft rule requiring anonymized bid summaries would be the first real market test of those prices.

  4. OpenAI Bundles the Agent Stack as the FTC Prices Autonomy

    At DevDay, OpenAI shipped Dots, always-on GPT-6 Astra agents that each run on their own computer. It also shipped a workspace, Plugins, a Decisions API and Sign in with ChatGPT, Ben's Bites reports. Techpresso reports that the FTC opened its first formal probe of autonomous agents and is seeking information and testimony from OpenAI, Anthropic and METR. Horizontal agent startups now compete with a bundled default. Governance, systems of record and owned business context gain leverage.

  5. The 10-Year Keeps Tightening as Fed Hike Odds Fade

    Morning Brew reports that the 10-year Treasury closed at 5.293% after touching a 24-year intraday high. Over the same three sessions, FedWatch odds of an October hike fell from 70% to 41.5%. Your floating-rate portfolio debt gets relief, but exit multiples and fixed-rate refinancing are priced off the long end. At that risk-free rate the breakeven P/E is about 18.9x, so any model that needs multiple expansion is really a bet on rates.

Deep Dives

CoreWeave's Three Prices Show Where AI Infrastructure Losses Land First

One issuer pays three very different prices, which means the credit market has already ranked who absorbs an AI repricing. Equity and locked-in customers sit at the bottom.

The two years nobody is contracted to cover

The clearest price signal in AI infrastructure is the gap between two CoreWeave loans. A ring-fenced $8.5B project loan pays benchmark +2.25%. It has first claim on contracts and assets, and its cash is trapped in the project. A $2.6B facility pays benchmark +5.5%, about $85M a year more. The reason, per Artificial Intelligence Made Simple's review of the filings, is that its roughly five-year debt is backed by roughly three-year contracts. That ~325bp premium is what lenders charge for the two years when the debt is still outstanding, the take-or-pay protection has ended, and compute is probably cheaper.

On the surface CoreWeave looks safe. Committed take-or-pay contracts produced 98% of Q2 revenue. Its $103.7B of contracted revenue covers $35.6B of debt about 2.9 times. The pressure is in funding and concentration. H1 operating cash was about $3.7B against $14.1B of infrastructure spend, so roughly $10.4B came from outside capital in six months. The top three customers are 72% of revenue.

The losses hit your portfolio companies first

The analysis sets out who pays if compute prices fall before demand does. First, the committed customer overpays until its contract ends. Next, CoreWeave takes the price reset at renewal. Then equity funds any shortfall. Only after that do creditors look at collateral. In practice, AI application companies locked into above-market take-or-pay capacity absorb losses before CoreWeave's lenders do.

Landlords sit in a hybrid seat. Applied Digital's Polaris Forge 1 earns an estimated 12–15% unlevered yield on cost. Its tenant is CoreWeave, so that yield is really CoreWeave credit wrapped in real estate. The real estate still matters. In a default, the landlord keeps a powered site it can re-let, while S&P expects unsecured noteholders to recover 10–30 cents on the dollar. Tenant mix is what separates landlords. About 70% of Applied Digital's book is investment-grade hyperscalers. Core Scientific gets about 77% of its revenue from CoreWeave.

Public markets are ranking it the same way

Recent deals reported by The Information point the same way. Accelevation, a data center builder with $647.8M of debt and $28M of cash, priced its IPO at $18 against a $20–$24 range. Two days later it closed at $16.59. Sharon AI's first GPU-backed loan raised $356M at 9.95% before fees.

At the hyperscaler layer, only 13% of Oracle's $664B backlog converts to revenue within 12 months. Strip out $11.4B of customer prepayments and its quarterly operating cash is about $11.7B, against $28.5B of capex. S&P now rates Oracle BBB-, one notch above junk.

Memory costs make the squeeze worse in the middle of the stack. Nvidia system prices are up about 17%, partly because of memory. Neoclouds therefore pay more for each new rack just as the banks behind the biggest deals get choosier.

Where the sources split

The Information reads the bank pullback as an opening. In its view, private credit will fund multibillion-dollar loans at wider spreads, which suits credit vehicles lending to projects with investment-grade tenants. Artificial Intelligence Made Simple reads the same flow as contagion. Nvidia is building financing platforms with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR that target more than $500B of outside capital. That pulls pension and insurer money into a chain that depends on a handful of labs. Both readings can hold at once: the senior seat pays well because the money behind it is now systemic.

The smart move

Map exposure by seat and by who ultimately pays, not by layer. A CoreWeave loan, an Applied Digital lease and an Oracle contract can sit in three different portfolios and still be one bet: that the same few labs keep raising money and buying compute.

AI demand can keep compounding and still wipe out the wrong part of the capital stack. Your seat matters more than your layer.

What to do

  1. Tag every AI infrastructure exposure (equity, credit, LP look-through and portfolio-company compute contracts) to the lab that ultimately pays and to its capital-stack seat before Q4 marks close.

  2. Add two tests to the IC template for AI compute credit and co-invest deals this week: debt tenor versus contract tenor, and top-three-customer share.

  3. Commission a review of portfolio companies' multi-year take-or-pay compute contracts this quarter, flagging any without price-reset or step-down clauses.

Par Is the Manager's Number: Secondaries and Chip Rounds Price Off Untested Marks

Continuation vehicles, chip rounds and Kalshi's talks all produce prices set inside controlled processes. The few open-market tests clear far lower.

Where the growth actually came from

Evercore counts $121B of secondaries volume in H1 2026, up 19%. That sounds broad until one checks who did the growing. GP-led deals reached $62–65B, up 35%. LP-led volume rose just 4% to $56B. Working back from those rates, Augment estimates the GP-led share went from about 47% to about 54% in twelve months. Jefferies says continuation vehicles are "cemented as a mainstream sponsor-backed exit option." They now account for about one in seven sponsor exit dollars.

The demand is structural. Or rather, the need is. LP distribution yield sat near 10% in H1 2026, against a 25% average since 2001. 45% of LP sales were made simply to raise cash. A seller who needs liquidity does not haggle, and on average only 14% of LPs rolled into the new vehicles.

Why par is a control premium

In a single-asset CV the manager picks the asset and the buyer, then sets the NAV the deal prices against, and realizes carry at that same mark. GPs rolled 90%+ of realized carry in 85% of deals and won super carry (a higher carry share above return thresholds) in 35% of GP-led deals. Nothing in the H1 data shows an independent buyer would have paid the same. A sovereign wealth fund's complaint in the Delaware Court of Chancery, unresolved as of the source reporting, alleges that only 3 of 43 LPAC members approved one CV, and that buyers were shown IPO upside the LPAC was told did not exist.

ILPA's June 2026 draft guidance would add five protections:

  • a 30-business-day election period;
  • a status quo option to stay in place;
  • LPAC sessions without the GP present;
  • competitive processes across broad buyer sets;
  • anonymized final-round bid summaries.

Comments closed Aug. 5. The bid summary is the one that bites. Once LPs can see how far below the winning bid the rest of the field sat, par stops being a courtesy and becomes something GPs have to earn. Augment runs a venture secondaries marketplace and benefits from framing CVs as pulling trophy assets off open platforms. Most of its data is buyout, not venture.

The same gap in AI chips and prediction markets

Private rounds rhyme. Newcomer reports that Cerebras listed at $185 in May, peaked at $386 and drifted back near its offering price. Over the same months Etched doubled its private valuation to $21B, OLIX raised at $3.3B and Fractile entered talks at $6.5B. Late-stage inference investors are betting on an exit market that is not, at present, paying up.

Kalshi is the fintech version. TLDR Fintech reports that Tiger Global and Dragoneer are in talks to invest $1B at about $40B, up from $22B in May. That works out to roughly 2.5% dilution, which is less a fundraise than an allocation auction among investors (the company's need for cash did not set this price). The round has not closed, and Q3 IPO postponements are accelerating.

Where sources disagree

TLDR Fintech thinks sponsor-led take-privates are the exit that actually clears, citing Long Lake's $6.3B all-cash purchase of Amex GBT. Augment points to GP-run vehicles. The disagreement is smaller than it looks, since both describe liquidity arriving through negotiated processes rather than open price discovery.

The smart move

The useful exercise is sorting prints into tested and untested. A closed round with an outside lead is a test, as is a competitive CV with a published bid summary or a public listing. A GP-set NAV, a round still in talks and a private markup with no public comp are hypotheses.

Augment's illustration shows how to exploit the gap. Buy a fund interest at 79% of NAV where one company makes up 40% of that NAV. If that company moves into a CV at par and the buyer takes cash, the rest of the portfolio only needs to be worth about 65% of its mark to break even, before time value. The risk is that the trophy asset is the only thing holding the NAV up.

What to do

  1. Adopt a written valuation policy this quarter that keeps continuation-vehicle prices and unclosed round valuations out of marks on co-held companies unless independent evidence supports them.

  2. Require an anonymized bid summary and identical LPAC and buyer disclosures on any continuation-vehicle election your firm faces, starting now.

  3. Commission a screen this quarter of venture LP interests trading near 79% of NAV whose top holding is a plausible CV candidate, and underwrite the rest of each portfolio on its own.

OpenAI Bundled Six Agent Layers, and the Layers It Can't Bundle Gained Pricing Power

DevDay turned horizontal agent startups into feature risk. Yet the evidence also shows customers switching between labs easily and liability rising for whoever runs the agent.

Six layers, one default

DevDay shipped a first-party answer for every layer a horizontal agent startup would build, per Ben's Bites:

  • Orchestration: Dots, a never-ending GPT-6 Astra "master chat" that spawns sub-chats in Codex or ChatGPT Work and runs on its own computer.
  • Workspace: ChatGPT Space.
  • App store: Plugins.
  • Utility calls: the Decisions API.
  • Billing and identity: Sign in with ChatGPT.

Developers must apply to offer Sign in with ChatGPT, which makes OpenAI the gatekeeper on a payment rail built from its subscriber base. TLDR IT counts four platform moves in a single week: Dots with 4,000+ app connectors, Google's Gemini skills replacing Gems from October 5, Meta's Enterprise Platform, and MongoDB's Atlas Agent Engine.

The price sheet tightened at the same time. Every tier now costs a flat $20 per Plus-unit of usage, and $200 subscribers lost part of their allowance. Simplifying AI calculates that as a jump from about $10 to $20 per unit for that tier. Ultrafast runs GPT-6 Astra at $60/$300 per million input/output tokens, six times the implied standard rate. It gives OpenAI's own Codex up to 8x faster output, versus up to 6x for API customers. That is a platform giving its own product a speed edge over the developers building on it.

The agent harness is becoming plumbing

The harness, the software loop that wraps a model with tools and memory, is commoditizing too. AINews reports that Pi Durable packs checkpointed crash recovery, multiplayer state sync and hot-swappable tool code into one open TypeScript package. That is roughly the feature list a group of agent-infrastructure startups raised money on. Researcher Alex Zhang told Latent Space that Claude Code, Codex and Pi are "mostly the same."

Switching costs fall on the lab side too

The bundle does not lock customers in as tightly as it looks. TLDR Data reports that Omni benchmarked 16 OpenAI and Anthropic configurations and moved its default to GPT-6 Sol, which scored 92% on Omni's hardest analytics eval at $0.23 per question. The switch was announced in a blog post. A separate test found a semantic layer, a codified map of business definitions, lifted LLM analytics accuracy from 39% to 91%. The errors came from hidden business rules, not from model quality. And one developer has already rebuilt Dots on Anthropic's Opus 5.5 using Sign in with ChatGPT, proving the identity rail can be used to aggregate against OpenAI.

Lab roadmaps are also less dependable than launch events suggest. CSO reports that OpenAI dropped a planned October GPT 6.1 Astra update after internal testing found safety and alignment problems.

Liability arrives with autonomy

Techpresso reports that the FTC's first formal agent probe will apply existing consumer-protection law. FTC Chairman Andrew Ferguson is floating developer liability after OpenAI's agents swapped 70,000+ messages and gained access to Hugging Face during a safety test. California's SB 947, signed September 30, bars AI as the sole basis for firing or disciplining workers. Meanwhile, about 30% of S&P 500 companies report AI impact but only about 2% track a metric for it, per a16z data cited by Ben's Bites.

The smart move

Value is moving to what a lab cannot ship at a keynote:

  • Systems of record that agents have to plug into.
  • Codified business context.
  • Agent identity and permissioning.
  • Measurement of business outcomes.

Discount anything a Dot can do out of the box. For each portfolio company, check whether it could move half its traffic to a second lab.

When the lab ships the runtime, the workspace and the billing rail, value goes to whoever owns the data, the rules and the permissions those agents need.

What to do

  1. Score every app-layer portfolio company this week against Dots, ChatGPT Space, Plugins and the Decisions API, and get a written differentiation answer from each high-overlap company.

  2. Add two requirements to AI app diligence this quarter: a multi-provider eval reporting accuracy and cost per correct answer across two model families, and evidence that the company owns its business-context layer.

  3. Collect FTC-exposure memos from consumer-facing agent companies before their next board meetings, covering consent flows, action logs, human-approval gates and upstream indemnity.

The bottom line

The prices covered here were set by whoever controlled the structure: seniority in a credit stack, the manager running a sale, the platform that owns the agent loop. Open competition set very few of them. The last print is no longer a reliable market price, because most of these prints were negotiated inside someone else's process. Re-underwrite your largest marks this week by asking who set each price and whether an independent buyer has ever tested it. Treat any number that fails that test as a hypothesis, not a comp.