The Investor

The sequencing is the tell

Watch the order in which incumbents moved. Workday, HubSpot, Microsoft and Amazon are running AI discounts and free trials before defections appear in reported numbers. Vendors with real switching costs raise price and absorb some logo loss; pre-emptive discounting is a confession of weak differentiation, and it compresses gross margin and net revenue retention simultaneously. The Information's reporting frames these cuts explicitly as retention moves — which is why a mark built on 2023-era net revenue retention plus contractual escalators is the stalest number in a software book.

The other half of the same playbook contradicts it: migrating those same customers onto metered pricing that raises effective prices. Pegasystems quantifies it. Monthly GitHub Copilot spend went from $20,000 to $260,000 after Microsoft shifted to usage-based billing — roughly $3.1M annualized against a total software budget of $10-15M, meaning one coding tool now consumes 21-31% of all software spend, up from about 2%. CIO David Vidoni did not churn. He installed per-employee spending caps and complained publicly that vendors offload cost optimization onto buyers.

What Microsoft did next matters more than the bill

Microsoft offered one million credits — worth roughly $10,000 — for a month of Copilot Cowork, and Vidoni plans to test Cowork in finance and marketing. A billing dispute became footprint expansion for five figures, with reporting suggesting other customers received materially richer incentives. Two consequences for your book: incumbent switching costs are cheaper to defend than the AI-native displacement narrative assumes, and every portfolio company carrying a Copilot line item should be demanding incentive parity before its next renewal window.

The cost side points the same direction

Two independent signals say unit-price deflation is not reaching gross margin. Token prices have fallen 75% while customer AI bills tripled, because agentic workloads consume tokens far faster than prices fall — that figure traces to a sponsored placement, so treat the precision loosely and the direction as corroborated elsewhere. And xAI shipped Grok 4.7 at exactly Grok 4.6's price with a larger base model and output self-verification. A model that checks its own work before returning it burns more tokens per completed job by construction: flat list price, higher total bill, and app-layer companies will switch without telling finance.

The cleanest version sits in voice. OpenAI's own GPT Voice engineers state plainly that full-duplex serving must be measured in concurrent sessions, not requests, because the model never idles — you rent GPU memory and compute for every second a human talks, thinks or pauses. Text application economics lean on idle-time amortization. Full-duplex has no idle time, so any voice position whose plan is built on per-request cost is overstating margin structurally, not marginally.

ARR is the wrong metric for metered AI revenue. Ask for month-over-month consumption retention, the share of accounts that have imposed internal caps, and revenue concentration in the top consuming decile.

The caveat that keeps this honest

The cohort is dozens of tech-forward companies averaging 2,400 employees — Snowflake, Datadog, Cloudflare, AMD among them — and OpenAI is both a Zip customer and a named beneficiary of the shift it measures. True broad-market AI wallet share is plausibly 2-4%. Use the 8% as a 12-24 month leading indicator, not as market-size evidence in an investment memo. What survives the caveat is still large: in the most sophisticated buying cohort in the economy, legacy software is a mid-single-digit growth business, and the revenue replacing it is consumption the buyer is already governing.

The most informative number is four years old

Avathon's January 2022 mark of $1.4B sat against roughly $15M of ARR the company still carried in 2025 — about 93x. It took the company 14 years to reach $50M of ARR. That reframes the live round from "hot AI asset" to "stale mark finally justified," and those are different underwriting exercises with different defensible prices.

Revenue quality decides the price, not the growth rate

Barrick pays a base platform fee plus fees tied to results achieved. So ">$50M ARR" is a blend of recurring and performance-contingent revenue, and the $10-20M annual range is a performance band, not a contracted floor. Every 10 points of contingent mix should cost the multiple roughly three to five turns. That single split moves the deal more than any growth figure in the file.

Concentration compounds it. Five accounts carry the business: Aramco Digital is the largest, Boeing is both a major customer and a 9% shareholder, Airbus is also a customer, the Department of Defense is the fourth, and Barrick is already about 10% of expected 2027 revenue — which implies 2027 revenue of roughly $100-200M. Boeing's dual role raises a related-party revenue question and a latent conflict with Airbus. Losing any single account is a 10-25% revenue hole landing directly on a fresh mark.

The repeatable part is the buyer, not the company

Barrick signed a five-year AI deal roughly three months before floating a minority stake in the unit that produces most of its gold. An operational contract became an equity-story asset ahead of a listing. That gives you a screen you can ship immediately: mining, energy, chemicals and industrials with announced IPOs, spins or minority floats inside twelve months are a buyer cohort with a board mandate, a hard deadline and low price sensitivity.

The sales cadence is now measurable too — pilot in December 2025, live at Nevada Gold Mines in Q2 2026, five-year contract signed September 2026, about nine months, with full rollout needing two more quarters. Use that as the default underwriting cadence for physical AI, and note where hardware sits by comparison. Caterpillar chose to partner with FieldAI rather than build autonomy in-house; Hyundai and Boston Dynamics will not put Atlas into plants until 2028; Humanoid remains at the Schaeffler pilot stage. The operations software layer has an 18-30 month window to become the system of record before embodied hardware reaches scale.

Two risks the growth rate hides

Real-time computer vision flagging worker safety violations at mine sites in the US, Saudi Arabia and the Dominican Republic is a labor-relations exposure, not a feature — it meets works councils the moment the platform touches Europe. And with China's internet regulator probing DeepSeek and Moonshot over alleged data leaks to Anthropic, US-origin model and data lineage is hardening into a procurement gate for defense and sovereign-energy buyers: a moat for vendors who have it, a disqualifier for those who don't.

Every financial figure here — ARR, contract size, raise size — traces to a person with direct knowledge speaking immediately ahead of a fundraise. Treat it as favorably framed until an account-level ARR bridge says otherwise.

Read the structures, not the valuations

The offerings clearing this window are liquidity events wearing the clothing of financings, a distinction that matters mainly to whoever ends up holding the paper. Of Oura's 50M shares, 36.5M are secondary, Forerunner is exiting a 9.3% stake in full, and $526.4M of roughly $567M in proceeds services RSU tax obligations, leaving about $6.2M of net proceeds for the company itself. Bamboo Insurance's book (35M shares at $18-20, up to $3.24B fully diluted, ~$700M of proceeds) is entirely existing holders, with CVC monetizing. The money funds employee tax bills and the exit door. Late-cycle books shaped like this have historically traded badly.

The pricing datapoint still stands

Oura is pricing ~$2.2B at $40-44 a share, about $14.1B of market value and roughly $15.6B fully diluted, on $1.2B of revenue for the nine months to June, up 74%, with net income of $60.8M against $1.6M a year earlier. Membership revenue more than doubled to $240M, about a fifth of the mix, and that recurring line is what earns the multiple. Note the basis difference before citing a number: $14.1B is market value and $15.6B is fully diluted — both are reported, and they are not in conflict. The part worth underwriting is the book: Eli Lilly taking up to $100M and Dragoneer up to $300M. Pharma capital is a direct bidder in consumer biometrics now, so the natural acquirer for a connected-health asset is a drugmaker.

The refusals are equally specific

Holtec pulled a $900M nuclear deal the night before pricing. SoftBank's SB Energy slipped from September to October seeking $5-7B at a contested ~$50B, against zero operating data center capacity. Nvidia's up-to-$3B anchor equity is 43-60% of the entire raise, and its $105B credit guarantee is roughly twice the equity value of the asset owner it backs. Strip the vendor capital out and genuine third-party demand for a $50B company is $2-4B.

The precedent inside the deal is more interesting than the delay, or rather it is the only part that outlives it. Nvidia bought $1.5B more of SB Energy at 90% of the eventual IPO price, taking its stake to $3B. The marginal price setter in AI infrastructure is therefore a strategic buyer with a negotiated 10% haircut, and anyone offered a pre-IPO infrastructure allocation on the same tape without that discount is subordinated by construction.

Two numbers explain the refusals. Standard Nuclear, listed in July, trades -20.6%; X-Energy, public since April, trades -36.7%. All of it is being marketed with the 10-year Treasury at 4.963%, and at that risk-free rate long-duration pre-profit equity is unfundable whatever sector label it wears.

Public books are funding demonstrated profitability near ten times revenue and refusing vendor-financed capex at any price.

The reporting agrees the sort is running on cash-flow profile rather than sector. It diverges on the metaphor: one read has the window "closing," another has it "bifurcating," and the Oura and Bamboo prints favor bifurcation, because non-AI issuers with real net income are getting done. This is probably wrong somewhere, and the two ways it breaks are worth naming: the window shuts on the profitable issuers too, or infrastructure reopens once vendor guarantees get priced honestly. The view anyway: for a late-stage book carrying both application-layer and infrastructure exposure, the mark-to-exit dislocation arrives from the infrastructure side first, while private application-layer pricing still holds.