ClarityBeta

Investment & Market Intelligence

The Investor

9 min read · 1,739 words · 3 moves to watch
The Signal

Situational Awareness was up 439% and still had to sell $10B to Citadel at a discount.

Nothing was wrong with the holdings. SanDisk, SK Hynix and CoreWeave all still trade above year-ago levels, and they bounced the day the selling stopped, which is the part worth sitting with: a roughly 30% drawdown on a levered book was enough to end the argument. Whatever private AI exposure you are carrying on the same financing terms does not get a next-day recovery.

In Play

  1. Middleware Exits at Absorption Value

    Anyscale, the commercial home of the Ray distributed-compute framework, sold to GPU cloud Nscale for $1.6B, per The Information; Bloomberg puts the price at $1.65B. Against a $1.38B September 2022 mark, that is roughly 16% appreciation over nearly four years — about 4% annualized — through the largest AI infrastructure buildout on record. Any orchestration, model-serving or MLOps position carried on 2023-24 comps now has a live exit print sitting below it.

    Ask Clarity
    Try

  2. Compute Rents Rise While Token Prices Fall

    a16z's data has H100 twelve-month contract rentals just under $2.50 per GPU-hour, nearly 40% above November; Kalshi forward-prices them at about $2.78. A100 spot is stable. Meanwhile OpenAI cut its low tier 80% to $0.20/$1.20 per million tokens and left the frontier tier unchanged. Input costs up, output prices down: any portfolio company monetizing the spread between rented compute and sold tokens is squeezed from both ends at once.

    Ask Clarity
    Try

  3. The First Forced Seller of AI Paper

    Situational Awareness, Leopold Aschenbrenner's two-year-old fund, ran a levered book to a reported $45B NAV, then moved over $10B of it to Citadel at a heavy discount, per Morning Brew. The Information reports the fund was up 439% through June before margin calls arrived. Nebius, Micron and SanDisk rebounded the next day, once the selling was done. That makes this a liquidity event — and the private positions on the same books have no mark-to-market rescue coming.

    Ask Clarity
    Try

  4. Corporate Balance Sheets Now Fund AI

    Corporate venture supplies almost 90% of dollars going into AI companies, against under 50% a decade ago, per Newcomer. Nvidia alone deployed $90B of corporate venture in 16 months across 283 rounds since 2021, committed $5B to Safe Superintelligence largely earmarked for its own hardware, and is reported in talks to guarantee up to $250B of OpenAI debt. The Batch sizes that guarantee at roughly 5.2% of Nvidia's $4.792T market cap. Your co-investors are increasingly buyers of demand, not returns.

    Ask Clarity
    Try

  5. Agent Identity Gets Its First Price

    Okta agreed to acquire Permiso Security for just under $200M, per TLDR IT. CyberScoop notes Okta disclosed no value, multiple or terms, and expects to close in Q3 of its fiscal 2027. Permiso's product watches employees, service accounts, applications and AI agents after they authenticate. Read it as a capability tuck-in rather than a revenue multiple: it caps entry pricing on standalone identity-detection deals while validating the agent-identity wedge sitting above them.

    Ask Clarity
    Try

Deep Dives

The Software Between the Model and the Metal Cleared at Cost of Capital

A capacity owner bought the orchestration layer, and the price it paid tells you which end of the AI stack collects the boom's returns.

What the buyer paid for

The price is in the map above. The buyer is the part worth sitting with. Nscale is roughly two years old and sells raw GPU capacity, which makes it a capacity owner rather than a platform company, and it has just set the clearing price for the commercial vehicle behind Ray, the distributed-compute framework used to schedule training and inference workloads across clusters. Ion Stoica's company was on The Information's Generative AI Takeover list. This was not a distressed asset being tidied up. It was a category leader exiting at cost of capital, in the middle of the most aggressive infrastructure buildout in the history of computing.

The two readings of that print do not conflict, and holding both is the useful position. Bloomberg's framing is bullish on the layer: a raw capacity provider moving up the stack turns scheduling into a contested asset and establishes neoclouds as credible strategic buyers of AI compute software. The Information's framing is bearish on standalone value: middleware between the model and the metal has no independent terminal value, and the exit clears wherever a compute owner decides it clears. Both are consistent with the same conclusion: the layer sells, but the acquirer sets the price, and the acquirer is whoever owns the hardware.

Why the marks are the exposed part

Most orchestration, model-serving and MLOps positions in venture books were priced in 2023 or 2024 against a thesis that inference volume would compound into platform economics. a16z's data explains why it did not. GPU rents rose while token prices fell, so the layer that monetizes the difference between rented compute and delivered tokens has been squeezed from both directions at once. There was never margin for it to grow into.

Position typeOld underwritingWhat the print says
Orchestration / schedulingPlatform economics on inference volumeStrategic absorption at roughly the last mark
Model serving / inference-as-a-serviceGross margin expands as tokens deflateSpread compresses from both ends
MLOps toolingIndependent scale outcomeBuyer is a capacity owner, on its terms
The infrastructure boom paid the people who owned the machines and the people who owned the workflow. It did not pay the software in between.

The smart move

The re-mark is arithmetic, and it is better done in-house than by an auditor in Q3. Anything in this layer carried meaningfully above its last round now has a public comp arguing against it, and the honest terminal case is a strategic absorption rather than an independent scale outcome. There is also a sequencing question with a short fuse: compute owners are acquisitive because they are raising capital against growth, so that appetite tracks the financing window rather than strategy. The buyer set that exists today is not guaranteed to exist after the next credit repricing.

The second-order read matters more than the markdown, and this is the part that is probably wrong at the edges but right in shape. If capacity owners are the natural acquirers of the software that schedules capacity, then the capacity layer and the workflow-owning application layer are the two ends of the barbell, and everything between them is underwriting a sale rather than a franchise. That belongs in the thesis document as a stated position, not as an inference each partner draws privately.

What to do

  1. Re-mark every orchestration, model-serving and MLOps position against the $1.6B-versus-$1.38B comp before the Q3 valuation committee, flagging anything carried above 1.2x its last round.

  2. Commission a written strategic-buyer map for each middleware holding this quarter, naming which compute owners could absorb it and what evidence exists that they are still acquiring.

Input Costs Up 40%, Output Prices Down 80%: Whose Margin Is It?

Two price curves moved in opposite directions this quarter, and which side of them a company sits on decides whether cheap tokens are a windfall or a write-down.

The two curves, stated plainly

The rental and forward prices sit in the map, so the interesting material is underneath them. A100 spot pricing is stable, which is the quiet part: prior-generation silicon is holding its rents, meaning the three-to-four-year depreciation schedules baked into neocloud and GPU-backed credit models are running conservative rather than heroic. Power conversion equipment is the one import category going the other way, with volumes about 23% below January 2025 and prices about 25% higher. That is an isolated bottleneck. It is not general inflation.

On the capital side, The Batch reports Meta placed $12.55B of data center bonds at a 7.5% coupon, 287.5 basis points over the ten-year Treasury and 50 basis points wider than its own earlier Louisiana project. Same issuer, worse price. There is a reading in which this is calendar and duration noise, and it is not impossible, but fifty basis points of widening on an identical credit does not usually mean the calendar. The marginal cost of AI infrastructure capital is repricing in public view.

On the revenue side, the cut was selective in a way the headline buries: alongside the 80% low-tier reduction, OpenAI took the mid tier down 20% to $2/$12 per million tokens and left the frontier tier untouched. AI Breakfast's read is that the gain came from serving-stack engineering, speculative decoding and context compaction, which is the difference between a promotional cut you can wait out and a quarterly cadence you have to underwrite.

Same headline, opposite trades

The distinction that matters for a book is not AI versus non-AI. It is who pockets the deflation and who eats it.

Revenue modelEffect of an 80% token cutEffect of 40% higher GPU rents
Token-metered resale / thin API markupASP collapses; needs multiples of volume to hold revenue flatFloor rises under a falling ceiling
Seat or outcome priced applicationPure COGS relief, unearned gross margin expansionAbsorbable if pricing power holds
Capacity owner / neocloudNeutralCaptures the spread; residuals beat schedule

The complication is that per-token deflation is not per-task deflation. Stripe's internal AI platform disclosed that most sessions require many turns, and multi-turn agents replay context on every one of them. Devshot reports Amazon booked a $1.8M bill on a routine coding task, 860% over budget, discovered afterward in usage metrics rather than beforehand in a limit. The Batch's harness data puts the same 70% task success at $8.23 per task versus $17.32 one model generation apart. Cost per completed task, not cost per token, is the only number that describes a gross margin.

A founder citing token price cuts as structural margin improvement is describing an input and calling it an outcome.

The smart move

Every AI position needs one of three tags: COGS-tailwind, ASP-headwind, or neutral. The exercise takes days, which is days not spent screening the next deal, and it is still the better use of the days, because two companies trading at the same multiple today sit on opposite sides of it. Then push a compute cost of $2.50 to $2.78 per GPU-hour through the models, rather than the declining curve most decks still assume, and identify which companies have committed-use contracts renewing inside two quarters. Serving-stack gains could outrun the power bottleneck, or the bottleneck could spread, or nothing moves and the tagging exercise was cheap insurance. Those renewals are where the surprise lands.

What to do

  1. Tag every AI position COGS-tailwind, ASP-headwind or neutral, using share of revenue that is token-metered pass-through versus seat or outcome priced.

  2. Re-run compute COGS across inference-heavy holdings at $2.50-$2.78 per GPU-hour and list every committed-use contract expiring within two quarters before the next board cycle.

  3. Require fully-loaded cost per completed task and median turns per session in every AI diligence pack, replacing cost per token.

The Thesis Was Right and the Fund Died Anyway

A levered AI book liquidated into Citadel while its holdings still traded above year-ago levels, which makes structure — not adoption — the risk your reserve models are missing.

The mechanism, not the drawdown

Newcomer's account is the one worth carrying into the partners meeting. Situational Awareness's core holdings, SanDisk, SK Hynix and CoreWeave, are still trading far above year-ago levels. A roughly 30% drawdown from peak ended the fund anyway, because the positions were financed. Morning Brew supplies the timing detail that turns anecdote into diligence: Nebius, Micron and SanDisk all rebounded the day after the block cleared. The seller sold the bottom because the seller did not pick the date.

The LP roster is the transmission mechanism, or rather the more interesting version of it. The Information lists Jane Street, Meta executives, Nat Friedman, Daniel Gross and the Collisons. When capital of that quality becomes a forced seller, the paper is not marketed; it is placed. Newcomer also notes that on Caplight, single-layer SPVs have overtaken direct-to-cap-table trades in secondary volume this year. Fee-layered access vehicles with thin information now dominate the flow you would be buying into.

Where sources disagree

The firm disputes reports that it shopped its Anthropic stake. AI Breakfast reads the pattern differently, and the disagreement is load-bearing: the fund sold public equities outright while keeping a roughly $5B private Anthropic position. Read that way, a sophisticated AI-native allocator concluded that private frontier exposure beats public AI beta at current prices, which is the opposite of the panic interpretation. That reading is the less popular one and it may well be wrong. Both readings agree on one fact: liquidity, not fundamentals, set the clearing price.

The structural half nobody underwrote

The same Newcomer reporting is why this is a category risk rather than one manager's error. Corporate balance sheets now supply the dominant share of AI venture dollars, and the reported Nvidia loan guarantee stands against OpenAI debt of up to $500B. The Batch sizes that against a $4.792T market cap and an $852B private valuation. Nvidia's defense is that the deals are hardware-collateralized. Collateral comforts in a boom and rounds to a fraction in a bust. Vendor financing is what broke Lucent and much of telecom equipment after the dot-com unwind.

The symmetry cuts both ways on the public tape. Google booked a $99B second-quarter gain on equity investments, driven mainly by Anthropic and SpaceX marks. The first large reversal on a mega-cap's investment line becomes the markdown catalyst for the entire private complex.

The biggest risk in this market is not being wrong about AI. It is being levered while correct.

The smart move

Two exercises, both cheap, and both competing for the analyst hours that were going to go to new-deal screening. First, a leverage and counterparty sweep across the book: every company with equipment financing, GPU-collateralized debt, venture debt covenants, or a strategic investor whose money is tied to hardware purchase commitments. Second, a circular-revenue line in the IC memo template quantifying what share of a target's revenue comes from customers capitalized by its own strategic investors. Haircut that number in the model. Do not footnote it.

What to do

  1. Run a leverage and counterparty sweep across the entire book, flagging any company with GPU-collateralized debt, equipment financing or venture debt covenants and under twelve months of unrestricted runway.

  2. Rebuild reserve models this quarter assuming strategic and corporate co-investors do not participate in the next round, raising reserve ratios where corporates supplied more than half the last round's dollars.

  3. Add a circular-revenue disclosure line to the IC memo template before the next committee, quantifying revenue sourced from customers capitalized by the target's own strategic investors.

The bottom line

Every repricing in these reports ran through the capital structure, not the technology: adoption held up wherever it was measured, and the losses landed on whoever was renting a spread, borrowing against collateral, or sitting between the two ends of the stack. That breaks the working assumption that being directionally right about AI protects a mark — position in the stack and cost of financing now explain more of the return dispersion than conviction does. Commission one uniform re-underwrite across the book that answers a single question per holding: does it own capacity, own a workflow, or merely rent both?