ClarityBeta

Investment & Market Intelligence

The Investor

8 min read · 1,673 words · 3 moves to watch
SATURDAY, SEPTEMBER 5, 2026
The Signal

Oura's only cash-cleared price is $39.07 a share, set by its own $1.09B buyback.

The transaction that retired 27.9M preferred shares is also what produced the headline $924.3M loss, booked as a non-cash deemed dividend sitting on top of $60.8M of actual net income. Which matters if you are carrying a comp built off the $11B October mark or the $16B-plus figure weighed in May, because both quote a price nobody paid in cash. Worth resolving before either number goes into a memo.

In Play

  1. Oura's Loss Is an Accounting Artifact

    Oura's S-1 reports a $924.3M net loss attributable to stockholders, but the charge is a non-cash deemed dividend created when the company spent $1.09B of its own cash repurchasing 27.9M preferred shares. PitchBook counts $60.8M of net income underneath. For you, that repurchase is the only cash-cleared Oura price in existence, roughly $39.07 a share, and it sits far below the reported $11B October mark and the $16B-plus Bloomberg said the company weighed in May.

  2. AI Revenue Is Narrower and Shorter-Dated Than Its Multiple

    Ramp's spend data shows 1% of OpenAI and Anthropic customers generating 80% of their enterprise revenue, and those top spenders skew toward tech and AI product companies. Separately, Madrona found 77% of enterprises re-evaluate AI vendors every six months or continuously. Together they describe revenue that is both concentrated in venture-funded buyers and re-decided twice a year. Ramp's base skews startup and SMB, so treat the concentration as directional rather than exact.

  3. Neocloud Marks Rest on Vendor-Funded Backlog

    Figure committed $3.5B to Nscale for up to 100,000 Vera Rubin GPUs with stated intent past $6B, and Nscale is taking equity in Figure as part of the same deal. Fluidstack went from $7.5B in July to above $18B on a reported $50B Anthropic capacity agreement, with Jane Street leading the round while remaining the largest reported customer of rival Crusoe. Backlog struck this way is counterparty credit and delivery risk, not independent end demand.

  4. Frontier Prices Hold While the Substitute Tier Collapses

    OpenAI's GPT-6 Astra shipped at $10 per million input and $50 per million output tokens, 2.5x its predecessor and identical to Anthropic's Fable 5.1, per The Information's reporting. Underneath, Z.ai's MIT-licensed GLM-5.3-Flash scored 63% on the DeepSWE coding benchmark at $0.24 per task against Claude Opus 5's 74% at $11.84. Cost deflation reversed at the top and accelerated at the bottom, so a portfolio company without workload routing is structurally short its own gross margin.

  5. Siting Replaces Silicon as the Infrastructure Constraint

    Organized local opposition killed a $100B data center project backed by two of the world's largest private equity firms, and the defeat is being packaged as a replicable blueprint. Texas Republicans are breaking from the administration's pro-AI posture over data-center siting, in historically the most permissive US jurisdiction. Against that, a16z data attributes roughly 300,000 construction and trade jobs to the buildout since 2022, giving these projects a blue-collar constituency that cuts the other way.

Deep Dives

Oura's Only Executable Price Is a Buyback, Not a Mark

The headline loss in the filing is a balance-sheet artifact; the underwritable inputs are a cash repurchase price and an 89%-margin membership line that has to keep compounding to justify either reported mark.

The mechanics behind the headline loss

A deemed dividend is the entry GAAP demands when a company buys back preferred stock above carrying value: the excess is charged against income attributable to common holders, operations never see it, and the only thing that actually leaves the building is the repurchase cash. Oura spent $1.09B of balance-sheet cash over nine months retiring 27.9M preferred shares held from Seed through Series C-1, roughly 17% of preferred outstanding, and paid above carrying value. The excess is the $924.3M. Read as an operating result it looks catastrophic. Read correctly it is a liquidity event funded off the balance sheet rather than through a third-party tender, which is a duller sentence and a more accurate one.

The part worth caring about is that it was cash-settled, rarer in this asset class than any mark. $1.09B across 27.9M shares works out to about $39.07 a share. Forerunner Ventures separately realized $65M on 1.6M Series B shares, roughly $40.63. Two independent prints, both cleared in cash, four percent apart.

What the business underneath supports

Revenue for the nine months to June 30 was $1.21B, up from $697.6M, on 3.1M rings sold against 1.8M, with $262M of free cash flow. The valuation case is not the ring — or rather, the more interesting version of it isn't: membership revenue hit $240.5M, up 121% at an 89% gross margin, paying members doubling to 5M. Annualize it and there is roughly $321M of software-margin recurring revenue sitting on roughly $1.29B of annualized hardware, which on disciplined multiples supports something around $7-8B. Against that: $84.4M of fiscal 2025 warranty expense tied to Ring 4 battery defects, and no price range in the filing.

That gap is the whole argument. The reported $11B October mark needs the subscription flywheel to keep compounding. The $16B-plus level Bloomberg reported the company weighing in May needs it to compound at triple digits for years.

Where the reads diverge

There are three ways this plays out and the coverage has already split along them. One read treats "soaring revenue and widening losses" as a story that works in a risk-on tape and evaporates in a risk-off one, and notes the S-1 disclosure is worth more to Oura's wearables competitors than to Oura. A second is structural: two Nasdaq filings in one week, Oura alongside Olympus-backed Accelevation at $727M of sales, reopening the hardware and health listing window only for revenue-scale assets. A third holds this is the first genuinely cash-generative consumer hardware-plus-subscription comp in years, reopening a category effectively unfundable since 2022. This is probably wrong, but the structural read is the one that binds.

Two cash-cleared prints four percent apart are harder evidence than any paper mark, and they are the only Oura prices that actually exist.

All three agree on the consequence: the comp set health-hardware marks get measured against is being written this month, and it likely lands below 2021 private pricing. The private repricing happens between the S-1 and the pricing. After that, the work is off a public multiple.

What to do

  1. Rebuild the Oura comp off the $39.07 cash print and a sum-of-parts on the 89%-margin membership line, and set a written walk-away number before the amended filing discloses an underwriter range

  2. Reopen consumer health-hardware passes from the last 24 months and re-engage the two strongest before the listing prices

  3. Pull the Accelevation filing and build a data-center power and electrical comp model this quarter, since it is the first public multiple anchor for that category

Your AI ARR Re-Decides Itself Every Six Months

Two independent datasets change what an AI logo is worth: one names whose balance sheet the revenue comes from, the other times how long it stays.

The mechanism, not the statistic

The ARR multiple framework rests on one unstated assumption: revenue booked persists unless a customer makes an active decision to leave. Continuous re-evaluation inverts that. Revenue now persists only through an active decision to stay, taken twice a year or more often. That is consumption revenue wearing a subscription contract, and it should not clear at subscription comps. The practical translation for your model is a base-case gross revenue retention of 75-85% instead of 90-95%, stressed at 70%, with cohort renewal at month six and month twelve replacing logo count as the primary diligence metric.

Three things compound the problem

First, the pricing architecture is being rebuilt in public. Greg Brockman's line — "pricing tokens doesn't make any sense… the price per task is what matters" — is not rhetoric: OpenAI has begun letting some customers pay only when the model works, and Salesforce is overhauling how it charges for AI. That converts a slice of reported AI "ARR" into success-contingent revenue, which no standard diligence template currently isolates.

Second, supply is a single point of failure. OpenAI cut off Cursor's access to its models. If the most commercially successful company in the agent layer can be de-platformed by its supplier, so can anything in your book. In the same period, ChatGPT, Claude, Gemini and Grok degraded simultaneously, which Mashable flagged as unusual and pointing at a shared upstream dependency. Every portfolio company that told you a multi-provider diversification story had it invalidated in a single afternoon.

Third, some of the demand is related. Nvidia is reported to be funding roughly $2.5B of Thinking Machines' $5-6B round, and the whales driving lab revenue skew toward venture-funded technology buyers. You can be exposed to the same dollar twice: once as a direct holder, once as a customer's funding source.

Growth rate is commoditized information. Revenue provenance and revenue durability are not, and neither is priced today.

The bifurcation is the opportunity

Sources converge on the diagnosis and diverge usefully on the conclusion. The concentration data argues for a revenue-quality discount wherever AI-native buyers dominate the customer base. The retention data argues something sharper: if most enterprise buyers churn-test continuously, then any company that can evidence three consecutive re-evaluation cycles survived is demonstrating something genuinely scarce, and should command a widening premium as the market absorbs this. Both readings can be right, and the fund that can tell the two apart in diligence is the one that gets paid.

One honest caveat: the concentration sample skews toward startups and small businesses, which almost certainly overstates the tech-buyer share relative to the labs' true global revenue mix. Direction is what matters here, and direction is enough to change the questions you ask.

What to do

  1. Re-cut every AI position's ARR by end-customer type and flag any company with more than 30% of revenue from AI-native or venture-funded buyers, plus top-ten customer concentration

  2. Rewrite the AI diligence template to require buyer-side reference calls on re-evaluation cadence, the share of stated ARR that is outcome-contingent, and a tested time-to-switch off the primary model provider

  3. Publish revised AI multiple bands to the investment committee before the next new-deal vote, modelled at 75-85% gross retention

When the Compute Vendor Funds the Order Book

Three of this cycle's largest infrastructure step-ups price contracted revenue from counterparties the provider funds, invests in, or competes with — and the one attempt to actually clear a private book got paused.

Read the structure, not the number

The Figure-Nscale paperwork is the cleanest template yet: a pre-revenue-scale humanoid company signs a multi-billion-dollar forward compute obligation to a provider that is, in the same transaction, taking equity in the customer. Deployment does not begin until H2 2027, in Barstow, Texas, which is a long while to carry a mark. The benign reading is that Nscale needs an anchor tenant to finance the site, Figure needs guaranteed supply, and equity is how the risk gets shared. The second reading, or rather the more interesting version, is that the contract exists largely so both parties can point at it. Both readings agree it is not an independent demand signal, and any mark treating neocloud contracted backlog as evidence of durable end demand is resting on thinner support than it was last week.

One shareholder, two competitors

The Fluidstack chain is more specific. Its step-up from $7.5B in July to above $18B rests on a reported $50B multiyear Anthropic capacity agreement, and Jane Street led the round while simultaneously being the largest reported customer of Crusoe, which repriced from above $10B in October to roughly $30B on a roughly $13B, five-year Jane Street contract. One firm now holds an information and pricing-leverage position on both sides of a competitive pair. Every other shareholder in either company inherits that conflict and is paid nothing for it.

Counterparty quality differs sharply as well. Crusoe's anchor demand is funded by trading revenue and hyperscaler balance sheets. Fluidstack's is funded by a company that has not filed an S-1 and whose federal standing is unsettled: a senior Defense Department official reaffirmed its supply-chain risk designation one day after the Commerce Secretary said the issues were resolved, and a week after a federal court found the Department had punished the company unlawfully. This read is probably too cautious, but Fluidstack equity is best understood as a levered position on that one balance sheet.

The funding base moved to credit

Underneath all of it, AI capex has migrated into the credit markets. ByteDance upsized its largest-ever syndicated loan from $20B to $30B, Asia's second-largest dollar loan of 2026 behind SoftBank's $40B bridge. Alibaba is placing roughly $10B of equity. DeepSeek is seeking $7.4B, with infrastructure spend running from ¥1.2B across all of 2025 to ¥11B in seven months of 2026, and every yuan of that is capital not deployed against anything with a shorter payback. Broadcom tripled AI revenue to $16.7B, guided $34.8B against $35.1B consensus, and lost about 4% after hours on a 0.9% miss. Sam Altman, the segment's largest counterparty, called the buildout "unsustainable silliness" while naming neoclouds.

When the compute provider is also the equity investor, the contract value is a marketing number, and the marks have not caught up.

The honest counterweight arrived the same week. PayPal paused the sale of its 80-plus company venture portfolio, Plaid included, at a $900M-$1B ask run through Jefferies, with Axios framing it as paused rather than abandoned, a distinction that carries real content. The book was marked at $900M-$1B and the bid was not there at carrying value, which is what happens when someone tries to actually clear a diversified private position. Reported values are rising on contracted and related-party revenue while executable values get tested and come up short.

What to do

  1. Run a related-party screen across all AI-infrastructure holdings and pipeline this quarter, quantifying what share of contracted backlog sits with customers the provider holds equity in, and haircut revenue quality accordingly

  2. Demand contract-level disclosure on the reported Anthropic capacity agreement — take-or-pay terms, prepayment, termination rights — before re-underwriting any Fluidstack-linked exposure

  3. Stand up a monitoring dashboard on Asian dollar-loan spreads and the reception of ByteDance's $30B syndication as a leading indicator for late-stage AI marks

The bottom line

The filings, rounds and survey data all separate two things the last two years let investors treat as one: value that was reported, and value someone paid cash for without help from the seller's own balance sheet or a supplier's equity. The assumption that breaks is that a headline round, a contracted backlog and a signed enterprise logo are the same grade of evidence — they now deserve three different discount rates, and the spread widens every time a filing replaces a press release. Commission a provenance review of every private mark you carry this quarter: name who funded the counterparty, and re-underwrite whatever you cannot answer for.