The Map and the Mark

The map

We have mapped thirty-seven late-stage private companies and recorded, for each, not the number alone but the event that produced it. The aggregate stated value is $3.48tn. Below the two largest AI laboratories, Anthropic and OpenAI, the map falls away sharply. ByteDance stands at $550bn, then a gap of $360bn to Databricks. The median company here is worth $13.3bn. Artificial intelligence and its adjacent categories account for fifteen of the thirty-seven names and 63% of the value. Aerospace and defence is the second-largest cluster by count, with five names and 6% of the value; fintech and markets carries six.

The table below carries the thirty-three names whose most recent mark rests on a specific, publicly disclosed transaction. Four further names are held to a later section, because the prices attached to them rest on an intention rather than an event.

The last three months

Nine marks on this map became public between 24 May and 24 August 2026. Anthropic’s $65bn Series H closed at the start of that window and is set aside for the reason given above. What remains is the more useful picture: where late-stage capital actually went this summer, and at what price.

The largest of those rounds was Prometheus. On 11 June the industrial-AI company co-led by Jeff Bezos and Vik Bajaj raised $12bn at approximately $41bn, from Bezos himself alongside JPMorgan, BlackRock, Goldman Sachs, DST Global and ARCH, taking total funding above $18bn. Prometheus builds systems for designing and manufacturing physical objects, from jet engines to pharmaceutical compounds. It sells no model and has no consumer product. That single round is still 1.4 times every other closed primary disclosed on this map in the same three months, added together, Databricks’ $5bn close included.

Two of the eight repriced on operating evidence rather than on scarcity, and both are worth reading closely. Cognition AI raised more than $1bn on 27 May at $26bn, co-led by Lux Capital, General Catalyst and 8VC. Eight months earlier it had closed at $10.2bn. Over the same twelve months its annualised revenue moved from $37m to $492m, and the share of its own code committed by its agent Devin rose from 13% in December to 89%. Together AI raised $800m on 1 July at $8.3bn, led by Aramco Ventures: two and a half times its mark of sixteen months earlier, on annual bookings reported above $1.15bn.

The three most recent prints on the entire map are days old. On 10 August Cambridge Aerospace, a British air-defence manufacturer whose Skyhammer interceptor is already in delivery, raised $300m at $3.4bn led by DFJ Growth, 2.6 times the mark it set in April. On 12 August Lovable raised $400m at $13.3bn from Menlo Ventures and the EQT-managed Scaleup Europe Fund, with run-rate revenue reported approaching $600m and a valuation double its December level. On 13 August Databricks closed $5bn at $190bn, led by Coatue with Blackstone, MGX, T. Rowe Price and Sixth Street Growth, on an annualised revenue run rate above $7bn.

What has not closed matters as much as what has. Databricks signed its term sheet at $188bn on 16 July and closed four weeks later at $190bn: the only one of these to resolve. Blue Origin’s first external round, at $130bn pre-money, was reported on 8 July. Anduril, Crusoe, Polymarket, Harvey and Perplexity were all reported in discussions at higher marks across July and August — Anduril at roughly $100bn on 24 July, eleven weeks after closing at $61bn, and Perplexity above $30bn on 23 August, with Nvidia named as a prospective investor. None of those prices has been confirmed closed, and Blue Origin alone carries $130bn of this map’s stated value.

What sets a price

Four different events on this map are all reported, in ordinary usage, as a valuation. They are not the same thing, and the difference is not academic.

Twenty-nine of thirty-seven names, carrying $2.51tn, are priced by a closed primary round. Four names carrying $837bn are priced by a secondary sale or tender: ByteDance, Stripe, Revolut and Kraken. That is one dollar in every four across the whole map, set by shareholders selling rather than by a company raising. One name, Blue Origin at $130bn pre-money, is priced on a financing that has been reported but not confirmed closed.

The direction of the bias matters. Primary post-money figures include preference structure, so they flatter what a common share is worth. Secondary prints exclude it, so they read low against a primary headline for the same company on the same day. Reading the two as one number is the most common error in this market, and it runs in a consistent direction.

Where the map disagrees with itself

Nine companies on this map carry two live prices at once: the mark that has closed, and a higher mark reported to be under discussion. The gap between them is not small.

The median of that set is 1.7×. Crusoe is the widest at roughly three times: a closed Series E above $10bn in October 2025, secondary interest reported around $23.6bn, and talks reported at approximately $30bn in July 2026. Three prices, one company, one week. Fluidstack sits at 2.4 times, Physical Intelligence at 2.0, Kalshi at 1.8 against a Series F that closed only in May. The largest company in the set is Anduril, at 1.64 times: Reuters reported on 24 July that it was in talks at roughly $100bn, eleven weeks after a $5bn Series H closed at $61bn, and one structure under discussion would commit investors now to a second financing at a higher price within about a year. The company’s response is the cleanest statement of the problem this issue is about: no decisions have been made, and anyone claiming to know the terms is speculating.

Disagreement also appears inside single events. Apptronik’s $520m Series A-X of 11 February 2026 was reported at approximately $5.3bn post-money by one outlet, at $5bn by another and above $5.5bn elsewhere; the company’s own release stated no valuation at all. Perplexity’s Series E-6 is reported in a $21.2–23bn range, and in early May 2026 its shares were quoted on secondary platforms at $69.54 and $63.16, the lower of which represented roughly a 9% discount to the round price; on 23 August a further round above $30bn was reported to be in discussion, with Nvidia as a prospective investor, giving that one company a range, a discount and a reported mark all live at once. PsiQuantum’s confirmed mark is $7bn from September 2025; a $10.5bn figure attributed to May 2026 appears in one place and is not corroborated.

The direction of travel

Nine companies on the map have two confirmed marks on the public record: the closest thing this market has to a return series. Eight moved up. One moved down. The median interval between marks was six months and the median move was 1.7 times.

The upward moves cluster tightly. Cambridge Aerospace at 2.62 times, Cognition at 2.55 and Together AI at 2.52 sit within a tenth of one another, in three unrelated sectors. Below them, Lovable at 2.02, Polymarket at 1.67, Revolut at 1.53 and Databricks at 1.42. The intervals are short: Cambridge Aerospace reset in four months, and Kalshi’s reported $40bn discussion began about seven weeks after its Series F closed.

Kraken is the only visible down mark. Payward, its parent, raised $800m in late 2025 at $20bn; in April 2026 Deutsche Börse paid $200m for roughly 1.5% fully diluted, implying about $13.3bn. A 33% reduction, taken through a strategic secondary rather than a down round, and nearly invisible had the buyer not been a listed exchange with its own disclosure duties.

The clock on the mark

A reasonable objection to any map of private valuations is that the prices are stale. On this evidence that objection is largely wrong, and the real problem is a different one. The median mark here is five months old. Twenty-seven of the thirty-seven were set in 2026, and twenty-four within the past six months. Only two marks are a year old or more.

The problem is not age. It is that a private mark becomes public when somebody chooses to publish it. Fluidstack closed an $830m Series A at $7.5bn in January 2026 and disclosed it on 20 July. For six months, the best available public price for that company was, by construction, wrong. No amount of diligence on the public record would have found the right one.

Lightmatter is the outlier at twenty-two months, still carried at the $4.4bn set by its October 2024 Series D with no newer round publicly disclosed. Figure AI, at $39bn, is the largest mark on the map with no refresh in 2026. Neuralink illustrates a third failure mode: its last clearly disclosed post-money is $9.6bn from mid-2025, and the public record on a subsequent $650m Series E is inconsistent across outlets on both timing and price. That mark is not merely old. It is known to have been superseded by something nobody has published.