The Investor

The arithmetic that stays out of the IC deck

Rollins compounded earnings at roughly 12% a year while its multiple travelled between 31x and 89x over ten years — and fell from 60x to 31x in the last twelve months with no earnings deterioration. At 12% growth, earnings double every six years. A 50% multiple compression therefore cancels six full years of operating compounding. That is the entire hold period of a typical growth-equity position, deleted by sentiment rather than by performance. If a public compounder with two decades of consistency can trade at both ends of that band for the same business, the exit multiple in your model is not an estimate. It is a preference.

The dispersion is the instruction

The 2026 AI-disruption panic in enterprise software round-tripped in about two quarters, but unevenly. Netting each drawdown against its rebound gives the market's live verdict on which business models AI actually threatens:

A 27-point spread has opened between two groups that traded as one sector six months ago, with no sign of convergence. Software got its multiple back; labor-arbitrage services did not. If you are pricing a services, BPO or IT-consulting asset off a blended tech-services basket, you are importing software's forgiveness into an asset the market has explicitly withheld it from.

The correlation almost nobody models

Medpace fell from about $450 to below $300 in April 2025 — not because demand for clinical trials fell, but because its biotech customers could not finance themselves. The four named causes were high rates, a shut IPO window, venture capital rotating into AI, and policy uncertainty. EPS kept growing throughout, and the stock now trades near $600.

That is the transmission channel that does not appear in a churn cohort until it appears as a revenue miss. Every picks-and-shovels company in the book whose ideal customer is a venture-funded startup outside AI carries this exposure, and the capital rotation that caused it is the same rotation your own fund participated in.

Where the two readings meet

The tape's reaction to the hike sharpens the point. Cyclicals broke while long-duration growth held flat — the inverse of every prior tightening cycle. One reading is that AI is now the crowded defensive trade rather than the rate-sensitive one, which converts duration risk into crowding risk. The other reading, from the round-trip data, is that the market rejected the substitution thesis for systems of record and kept it for labor arbitrage.

Both readings agree on the mechanism that matters to you: earnings did not halve on the way down or double on the way up. The multiple did all the work in both directions, and it is the input you control least and stress-test least.

Sell-side targets on the software names followed the stock rather than led it — which means your own comparable-company marks are procyclical by construction.

One discipline note: the drawdown figures and the ten-year multiple band trace to a single data provider with unverified aggregate performance claims. Re-check both against a second source before either enters a committee model.

The curve is non-linear at the top

The per-task economics make the trade visible. Astra Max costs $3.94 per task against $1.03 for Sol xHigh — 3.8x the price for roughly a 4.7-point quality delta. Fable 5.1 Max runs $4.40 against $2.07 for Opus 5 High — 2.1x for about 3.5 points. Any application-layer company calling the newest model by default is donating gross margin upstream, and complexity-aware routing recovers a large share of it. That is now a diligence disqualifier rather than a roadmap nicety.

Why the bill grows even as prices fall

The serving physics explain the paradox. In long-conversation inference, the KV cache — the stored attention state for every token already in the conversation — can consume more GPU memory than the model weights themselves, and it scales multiplicatively: double the cached tokens and you roughly double the cache; double the active sequences and you double it again. Revenue in these products scales with tokens. Cost scales with tokens times concurrency times conversation length. Those are not the same curve, and almost no deck distinguishes them.

That is the mechanism underneath a sophisticated buyer being surprised, and underneath the response: a dedicated sub-budget created to ration access to the best model. The buyer invented the cost-governance category before any vendor sold it to them.

The counter-evidence, stated honestly

Three independent price events landed the same week and all point down:

• Gemini 3.8 Live undercuts GPT-Live 1 by 7x on like-for-like realtime multimodal work.
• Jev offers 5x cheaper inputs than 5.6 Luna with free output tokens, built on a non-generative architecture that emits probabilities over candidate answers rather than tokens.
• Gemma 4 12B is free and Apache 2.0, multimodal, running at roughly 15 tokens/second on a base 16GB M4 MacBook Air with Wi-Fi off.

Longer-run, GPT-3.5-class capability fell from $20 to $0.07 per million tokens in 23 months — roughly 280x. Both things are true at once, and the reconciliation is the investable part: the price of a fixed capability level collapses toward zero, while the cost of a task at the frontier rises, because tokens-per-task on long-horizon agentic work grows faster than per-token price falls.

The demand-side tell

Developers are now buying less output on purpose. One community tool benchmarks roughly 54% less generated code on average and up to 94% at peak across twelve tickets; another exists purely to cut token spend. When users engineer their own consumption down, usage-based pricing has a demand ceiling nobody has modeled — and agent spend governance has a hard-dollar ROI pitch with no funded incumbent.

Revenue scales with tokens; cost scales with tokens times concurrency times conversation length. Until you underwrite that gap you are paying for gross margin you have not seen.

Three caveats before this enters a memo. The per-task deltas are Arena-reported on a specific harness mix, not a benchmark. The 60% spend increase could be a migration artifact that normalizes within two quarters — watch for a follow-up disclosure. And one of the cheap frontier-class models has murky provenance and may be a router artifact, so treat it as a pricing signal rather than a capability fact.

Who can actually write the check

The March round listed nearly 30 investors — T. Rowe Price, Fidelity, BlackRock, Coatue, D1, Dragoneer, a16z, Sequoia — and almost none of them can write a cheque in the tens of billions. The list tells you who was allowed in. Look instead at who anchors, and the condition each one is in:

Strip out the levered holdco and the largest backers are the company's own chip and compute suppliers. The suppliers are buying equity with money the lab paid them for chips. When the marginal buyer of an asset is the seller's vendor, the valuation stops working as a price signal and starts working as an accounting convention.

The efficiency inversion the pricing has not absorbed

OpenAI's annualized revenue passed $40 billion in July against $182 billion raised or committed, which works out to roughly $0.22 of revenue per dollar of equity. Anthropic sits at $65 billion on $130 billion, or about $0.50. So the lab seeking something like 30x revenue is the less capital-efficient of the pair, and it is doing that next to a disclosed ~23x comparison at an estimated $1.5T listing.

Qualify each figure precisely before it goes anywhere near a memo. The $852B mark was a March financing-round valuation on a $122B raise. The $1.2T is investor-initiated early talks with no formal process reported, and the higher internal figure is stated belief. The $1.5T is an estimated IPO comparison. None of these can be traded on; the only completed transaction in the set is the ~$7B August employee tender, a secondary.

Why staying private is the strategy

The sharpest claim in the reporting is about why the IPO is deferred. Staying private buys time to grow sales without disclosing what the growth costs, and the cost in question is the margin-crushing discounts used to pull enterprise business away from the rival. Deferring the listing also forgoes the public buyer pool. That leaves the chip vendors as the only cheques big enough. And it makes the rival's prospectus the dated catalyst, because public margin disclosure ends the free-fire zone and becomes the discount-rate reference for every private AI mark on the book.

CoreWeave's edge is the open window

CoreWeave raised into two markets in a single day: an ATM of up to 35 million shares (roughly $3B at an implied ~$85.71 per share) plus $3B of convertibles with a $500M upsize option, which is $6.0-6.5 billion of potential capital in 24 hours, bought with real dilution and convert overhang. Set that beside a supplier extending up to $105B of credit support. The buildout is financed at the margin by open credit and equity windows, and those windows can close in a week.

The reporting splits two ways, and I'd take both halves seriously. One reading treats the semis selloff as a mispricing, since no capex plan changed and demand was reaffirmed in the same week. The other treats the funding architecture as the exposure. This is probably too tidy, but both can hold at once — demand intact, financing fragile — in which case the stress shows in spreads first and in revenue later. I'd expect the convertible to reprice the equity risk before the ATM does.

Compute constrains the buildout. Credit and equity market access constrains the compute. The $105B of credit support and the 24-hour raise both depend on windows staying open.

The underpriced tail: Bridgewater's Greg Jensen is arguing frontier labs should be regulated like systemically important banks. Capital requirements would break exactly the vendor-financing and circular supplier-equity mechanisms funding this cycle.