The Investor

What the acquirers were actually buying

The total is the boring number. The interesting one is whose treasury wrote the largest check: Palo Alto Networks — a security company — paid $3.35B for Chronosphere, more than Snowflake paid for Observe ($1B) and over three times Dynatrace's $915M for Arize AI. Order the prints that way and the category changes shape. Nobody paid up for IT monitoring here; they paid up for visibility and safeguards over autonomous agents, out of a security budget, on security multiples rather than APM ones. That $3.35B is also $3.35B not spent on anything else on the acquirer's roadmap this year, which is the part the press release never mentions. Dynatrace's Arize print sets a hard floor comp for agent and model-behavior tracing, and Elastic's $85M for Deductive AI sets the tuck-in floor underneath it. Call it an $85M–$915M underwriting band for seed and Series A in agent tracing, evals and guardrails. Printed in the last two weeks, not projected.

The mark dispersion the growth story left behind

Same category, same tailwind, same quarter, and roughly a 5x spread in implied multiple, driven almost entirely by when each company last went out for money.

Cribl's mark is 24 months old against ARR that has moved 33% in six months, and management has said out loud that it intends to list inside roughly two years. The repricing catalyst sits on a calendar rather than in a thesis, which is a materially different risk profile from waiting for a category to be discovered.

Where the workload is actually going

The quiet competitive datapoint is worth more than any of the deal prices: OpenAI's ClickHouse usage ramped sharply this year while OpenAI remains a Datadog customer. That is incumbent observability spend leaking to a cheaper, open-source-rooted alternative at the single most-watched reference account in software. The Information's Briefing and Dealmaker coverage corroborate the same figure from different angles, with ClickHouse past $350M recurring revenue and OpenAI and agent workloads named as the drivers. Cribl sells the identical arbitrage as arithmetic rather than vision: data compounding around 30% annually against flat budgets, executed for Zoom, ServiceNow and Hilton by routing telemetry into cheaper storage.

Four assets left the market in ninety days and a fifth publicly declined to sell — bidder demand now concentrates on a two-name shortlist.

What breaks the trade

Discount each number by its provenance, because the provenance is uneven: Cribl's ARR is self-reported, Grafana's "pickup" comes from its CMO with no figures attached, and the M&A-appetite color rests on unnamed bankers. This is probably the wrong worry, but the systemic risk is anchor failure: if Datadog's public multiple compresses on displacement evidence, every private mark in that table loses its reference point, and the 43x asset falls furthest. The counterweight is that this demand is cost-takeout driven rather than discretionary innovation spend, which makes it one of the few AI-adjacent categories that still works in a downturn. Both things can be true, and the calendar decides which one gets priced first.

Run the compression to the equity line

PitchBook's Kyle Walters supplies the ratio that does the work here, and it is a short one: assets "bought at 12x, and are maybe worth 10x." The instinct is to file that as a 17% problem and move on, which would be a mistake, because at 2021 leverage of roughly 5.5–6.0 turns of debt a 17% decline in enterprise value does not distribute itself politely across the capital structure. It lands on the equity strip, more or less entirely.

Before acting on any of that framing, the dull step: check the arithmetic. 33.8% of 13,509 companies is 4,566, which means the widely quoted "4,500 zombies" number is nothing more exotic than every company held five years or longer. The $860B is NAV in funds older than seven years — a different denominator, doing different work. Hold period is not distress. The genuinely abnormal cohort is the seven-to-ten-year bucket, and that is where the stranded equity actually sits.

Price discovery has effectively stopped

The transaction tape in Fortune's Term Sheet reporting is the evidence, and it is evidence mostly by omission. ATIS/AuditMate, Rotunda/Revv, and Linden taking ArtesRx off Flexpoint Ford all closed with terms undisclosed. Descartes paying roughly $100M for freight-TMS provider Tai is the only visible price in the batch. When nearly every sponsor transaction in a compressed-multiple environment closes dark, reported NAV becomes the least reliable number in the portfolio review.

The venture side of the same book shows the same symptom in a different instrument. Forge's July data broke a two-year pattern: the median secondary traded at a 7% discount to last round against par in June, and buy-side share of indications fell nine points to 48%, the first sub-majority month since late 2023. The honest read is dispersion rather than correction — 48% sits broadly in line with historical averages and comfortably above 2022's routine sub-40% prints — with a handful of AI names absorbing the available capital while the median private company quietly loses its bid.

A bid referenced off a GP's quarterly statement is a bid against a model, not against a market.

The resolution mechanism is already on the tape

Rotunda bought Revv to fold into AirPro Diagnostics. Thompson Street bolted AuditMate onto ATIS. Strong platforms absorbing aging sponsor-owned assets as add-ons is the trade being executed right now in ADAS repair software and elevator asset management, and seller leverage is at a structural low that is now public information. Which is the problem. The asymmetry disappears the moment these processes go competitive.

What would break it

The bear case is not the $860B. The bear case is that nothing forces resolution. GPs keep timing discretion absent a credit event, a recession, or a further rate move, and Fortune's own framing is that private equity is adept at kicking the can. Postponed loss recognition is not eliminated loss recognition, but it can comfortably outlast a fund's marketing cycle. And one month of one marketplace's composition-dependent median is not a macro thesis. September's print is the confirming datapoint.

The demand side is visible on the same page as the supply side

Hugging Face has the revenue shape that gets a company bought: more than $150M annualized, up 50% in two months, paying subscribers doubled in H1 2026, and a CEO willing to say "close to profitability" out loud. The reported price near $13B is roughly 87x, about 2.9x the company's $4.5B 2023 round, and The Information, which reported it, calls that steep even against recent AI M&A, naming Stripe's $7.5B OpenRouter purchase as the cheaper comp. Nothing has closed: this is a reported process, not an agreed price.

The other half of the trade sits in the same reporting. AT&T is deliberately routing workloads to open models to curb its Anthropic bills. Demand-side cause and supply-side beneficiary, printed the same day, which is tidier than these things usually arrive. Any portfolio company whose gross margin depends on reselling frontier tokens is on the wrong side of that precedent.

Enterprises had the option to move to the frontier and declined

Ramp's card index, built from spend across 70,000 companies rather than a survey, puts Anthropic's late-July flagship Opus 5 at 3.5% of the company's own model spend in July, against 28% for the older and cheaper Opus 4.8 and 8% for Fable 5. Anthropic's annualized revenue went from a reported $47B in May to $65B in July. Demand for AI is not the problem. Demand for the most expensive AI is. The counter-thesis is fair, or rather it is the version worth arguing: a model launched in late July sits early on its adoption curve, and July card data may measure rollout friction instead of preference. The 28% is what argues against it.

Before anyone runs a secondary, one disclosure discrepancy deserves diligence hours. Anthropic told investors it has 6,000 customers each paying $100,000+ per year. That is roughly $600M annualized, under 1% of the claimed run-rate. Either the base is overwhelmingly long-tail API and consumer revenue, far more elastic and churn-prone than a contracted enterprise book, or the headline number is doing heavy lifting.

And the methodology itself now leaks

In August 2026, researchers from MATS Research, the ELLIS Institute Tübingen and the Max Planck Institute for Intelligent Systems showed that the encrypted reasoning blocks OpenAI, Anthropic and Google hand back to API clients can be replayed into a cheaper sibling model in the same family, which prints the flagship's hidden chain of thought in plaintext. No cryptography broke. The authenticated field set covers model name and version but omits the account and conversation, so valid ciphertext works as a bearer token. Extracting frontier methodology rather than outputs runs roughly $720 per 10,000 traces, or zero when the blocks come from public datasets.

Flagship models get anti-distillation training; cheap tiers don't — so a model family's security collapses to its least protected member.

Two cautions, both real. The vendor matrix is a July–August snapshot and may already be patched, and the reconstructions are explicitly approximate; the weaker model is a fuzzy decoder, not a decryptor. This is probably wrong in the details, but what survives is structural. The assumed frontier-to-fast-follower lag shortens, and open-weight challengers re-rate upward on a relative basis. The labs still hold distribution, scale and enterprise trust. They hold rather less of a quality moat than most portfolio models assume.