ClarityBeta

Investment & Market Intelligence

The Investor

9 min read · 1,841 words · 3 moves to watch
WEDNESDAY, OCTOBER 14, 2026
The Signal

Disney hired Character.AI's CEO, then bid $20M for IP Google paid $2.7B to license.

The 2024 deal licensed the models and rehired the founders, and investors cashed out at signing, which was an excellent outcome for them and left nothing binding anyone to the company that stayed behind. So when the engineers followed the boss out the door, the shell was worth under 1% of the original fee. That is the real residual value of a license-and-rehire structure, and it is worth knowing before you hold whatever equity gets left behind.

In Play

  1. Acqui-Hire Floors Collapse

    Character.AI filed for Chapter 11 on Oct 12, and The Information reports Disney then bid $20M for an IP license. Google paid $2.7B to license the same company's models in 2024. Separately, Ithaca Holdings agreed to take Bumble private at $6.05 a share, about one-seventh of its $43 IPO price. Both deals show that a strategic buyer's past interest no longer puts a floor under consumer-AI paper.

  2. Physical AI Priced on Contracts

    The Information reports Applied Intuition has offers valuing it at $30B or more, roughly double its ~$15B June mark, and is likely to raise $1B+. Its only disclosed revenue is $400M of ARR at end-2024, so most of the step-up is a higher multiple, not growth. Runway separately says web video trained robots about as well as teleoperated data (16.1 cm vs 16.0 cm placement error). If that holds, signed deployments are a stronger moat than robot datasets.

  3. Enterprise AI Spend Loses Acceleration

    A Wharton/GBK study reported by The Information finds 45% of large U.S. firms will spend $10M+ on AI this year, up from 39% in 2025. Only 10% expect substantial increases next year, down from 19%. Only about 27% track revenue from customer-facing AI. For your enterprise-AI holdings, 2027 expansion now depends on customers who can prove growth outcomes, and Workday starts billing for agent actions in February 2027.

  4. Frontier Discounts Have Expiry Dates

    Artificial Analysis reports that Gemini 4 Argon's introductory price of $2/$10 per million tokens doubles to $4/$20 after Nov 30. Argon costs about 40% less per task than GPT-6 Astra. A straight-line estimate puts it about 20% above Astra once the price doubles. Meta's Muse Wonder is 60% cheaper per token than Fable-class models, but per task it is only about 24% cheaper than Opus 5.5 ($4.52 vs $5.98). Re-run any app-layer margin model built on launch pricing before December.

  5. Compute Control Moves to Megawatts

    The Information reports that Anthropic has signed $531B of compute deals in 11 months. It is also in final talks with NTT Data for a 500 MW Osaka site costing up to ¥1T ($6.3B). Microsoft's Maia 300 plan would need millions of chips, but Microsoft has produced only tens of thousands so far, so near-term relief on GPU prices is unlikely. For your infrastructure exposure, the open question is how much of these commitments revenue will actually cover.

Deep Dives

The License-and-Hire Exit Paid Once, and Left Nothing Under the Shell

Character.AI's investors were paid in 2024, so the bankruptcy judges the deal structure rather than their returns, and it tests what floor your downside cases still assume.

The order of events matters more than the size of the bid. Google's 2024 deal licensed Character.AI's models and rehired co-founders Noam Shazeer and Daniel De Freitas. The Information reports that investors cashed out at a decent profit. Karandeep Anand, hired as CEO in mid-2025, joined Disney as CTO in September, and other members of the technical team went with him. Character.AI filed in Delaware on Monday, listing assets and liabilities of $50M–$100M. Disney's bid came a day later. The talent left first, and the IP is being sold last.

The terms show what the asset is worth to a strategic buyer. Disney's bid covers a permanent but non-exclusive license to some IP and legal claims. Disney must also borrow back former staff to keep the service running safely. A buyer that doesn't need exclusivity is telling you the models aren't scarce. Without a buyer, the platform shuts down. Google paid $2.7B for its license and has since lost Shazeer to OpenAI.

The other 2024 license-and-hire deals ended in very different places. Adept's website has not been updated since mid-2024, while Inflection relaunched as Inflection AI Labs. In every case, the deal paid early holders at signing and left behind a company with no clear owner of its future.

Bumble: a distress price with an AI label

Ithaca Holdings is co-led by ex-Sequoia partner Ravi Gupta and ex-Meta CRO John Hegeman. It agreed Tuesday to buy Bumble and pitched the deal as an AI bet: Gupta said Bumble stood out among companies that "would become more valuable as AI advances." The numbers say otherwise. Revenue fell 10% last year and another 15% in H1. At June 30, Bumble carried $450M of debt against $154M of cash.

The two reports on price don't match. The Information's morning report put advanced talks at roughly $1B, about 3.3x Bumble's ~$300M market cap before the leak. Its later report puts the agreed deal at a $1.3B valuation. The ~$300M gap is close to Bumble's net debt of about $296M. That suggests one figure is equity value and the other is enterprise value (equity plus net debt), but neither report confirms it. Check the filing before using either number as a comp.

The holdco buyer class is real. Sequence Holdings, backed by 8VC, Lux and Conviction, closed a $7B take-private of an insurance broker with Michael Dell's DFO Management last month. That playbook depends on AI cutting labor costs in fragmented service businesses. A dating app with shrinking revenue is a different problem. At one-seventh of the IPO price, the entry discount is doing most of the work.

Cost-to-serve is the binding constraint

The distress traces back to unit economics. Users report using up $200 Muse plans in hours or days. The evidence is anecdotal, but it shows how heavy users break flat-rate pricing. Meta now names consumer use as Muse's main development focus. That puts a far better-funded rival into the exact segment where standalone consumer apps are failing.

In our view, the right move is in the downside case, not the upside. When a founder treats a hyperscaler licensing deal as protection, ask what the company is worth if no one calls. Then price the remaining IP the way the bid did.

A hyperscaler license-and-hire deal pays investors once. It is not a floor under whatever is left.

What to do

  1. Rebuild the downside case in every consumer-AI and model-lab memo before the next IC. Remove any hyperscaler license-and-hire floor and price the remaining IP at a recovery like Character.AI's.

  2. Commission a cost-to-serve review of every consumer-AI holding this quarter, comparing heavy-user inference cost with subscription revenue by cohort.

  3. Track the Character.AI docket through the expected late-November or early-December close and log the clearing price as a distressed AI-IP comp.

Physical AI Is Priced on Contracts Because the Data Moat Just Thinned

If ordinary web video trains robots as well as expensive teleop data, signed deployments become the defensible asset, and that is what the $30B offers reward and what most of the sector lacks.

How much of the step-up is growth

Applied Intuition reportedly doubles its revenue every year. Over the roughly four months since June, that compounds to about 26% revenue growth, while the valuation doubled. By that arithmetic, the revenue multiple itself rose about 1.6x, so most of the move is sentiment rather than fundamentals. The only disclosed revenue figure is $400M of ARR at end-2024, and against that, $30B is about 75x. Against an extrapolated end-2026 ARR near $1.6B, it is about 19x. The $400M figure and the doubling claim both come from an a16z source in July, and a16z is an investor in the company.

The company doesn't need the cash. It has raised $1.2B from a16z, General Catalyst, Lux and Kleiner and hasn't spent most of it. Raising $1B at $30B means only about 3.3% dilution. We read the raise as a way to get priced stock for acquisitions in simulation, synthetic data and validation tooling.

What buyers are paying for is breadth across automakers. Honda and Nissan signed last week, joining Isuzu, Volkswagen and Humain. The product now covers development and validation and extends into the vehicle operating system. That is contracted revenue over long automaker product cycles. It sets the company apart from the agent and search companies The Information places beside it in the $20B+ club.

Why contracts, not data, are the scarce asset

Two separate teams now claim that cheap video can stand in for expensive robot data. Runway pretrained robot policies on ordinary web video and on teleoperated robot video, 10 to 1,000 hours of each. It measured object-placement error of 16.1 cm for web video versus 16.0 cm for robot video. About 30 minutes of robot data then adapted its Praxis-1 model to a robot body it had never seen, and Runway plans to release the model weights openly. Dyna Robotics told Latent Space that its base model trains on more than 1M hours of human video and learns new tasks from as little as 45 minutes of demonstrations.

Both claims come from the companies themselves, and 16 cm of absolute error shows cheaper data, not solved manipulation. If the claims hold, though, a startup whose edge is a proprietary teleop dataset owns a depreciating asset. A deployment that earns renewals does not depreciate. Dyna says its robots fold thousands of napkins a day across many Din Tai Fung US sites, with 99.4% success over 24 hours. Din Tai Fung then signed a "much bigger contract" that adds food cutting. That is the robotics version of Applied Intuition's automaker wins. It still involves only tens of robots, and neither revenue nor contract value has been disclosed.

Where the evidence is weakest

Runway's growth figures appear in a sponsored essay in Not Boring co-written with its co-CEO, and they don't fit together. The company says it added more than $100M of ARR in Q3 2026 and grew about 50% month-on-month. Sustained 50% monthly growth compounds to about 3.4x in a quarter, which would imply a starting base near $42M. Bloomberg adds a pricing complication. DeepSeek, with more than $12B of financing, aims to lead the first round of a months-old robotics startup founded by an ex-Tencent scientist and tied to a $400M robot lab. When strategic capital sets first-round prices, financial investors struggle to win lead positions on good terms. The $400M figure is also ambiguous.

The practical consequence is two comp sets. One includes the sentiment-driven $30B print, and one is anchored to revenue. Deals priced between them are defensible. Deals priced at the sentiment level need contracts to back them up.

Physical AI with signed contracts now gets priced like frontier AI. The rest of the sector is still priced on demos.

What to do

  1. Rebuild the autonomy and robotics comp sheet this quarter. Show multiples on both Applied Intuition's disclosed end-2024 ARR and the extrapolated ARR, and commission a channel check on its current ARR.

  2. Add a deployment screen to every robotics diligence this quarter. Cover sites live, robots deployed, 24-hour reliability and contract expansion history. Ask what the company keeps if web-video pretraining plus ~30 minutes of robot data closes its gap.

  3. Commission a regulatory and cap-table review before co-investing alongside Chinese strategic capital in any embodied-AI round.

Enterprise AI Buyers Are Still Spending but Can't Prove Why, and Metering Arrives Next

Agent metering will put AI on the invoice just as buyers' plans to expand weaken, while banks and retailers restrict where outside agents can act at all.

Budgets measured by proxy

The Wharton/GBK survey of 1,400 U.S. employees, at firms with 1,000+ staff and more than $50M in revenue, is mostly a study of how buyers grade AI, and the answer is generously. Forty percent of managers score it on hours saved per employee, 37% on fewer errors. Study author Jeremy Korst calls the assessments more "vibes" and "perception" than hard numbers. The 28% reporting strongly positive returns cluster at companies that measure growth outcomes. Only 8% expect new revenue or products to be AI's main role over the next one to two years. Productivity gets 44%.

Then there is the seat count. Visier found that 48% of 1,000 U.S. workers exaggerate their AI usage or expertise at least some of the time, per Morning Brew, and 59% don't believe, or aren't sure, that their employer has an AI plan. If mandates drive purchases while real usage lags, seat-based ARR is overstated, and the renewal risk has been postponed rather than retired.

Proof of new revenue is rare, and what exists is company-reported. Walmart says shoppers using its Sparky assistant spend 40% more per order, which may say more about who picks the tool than about the tool. Novo Nordisk expects "tens or hundreds of millions" a year from AI agents that audit and coordinate clinical trials. That range spans an order of magnitude.

Metering puts the cost on the invoice

On Oct 13, Workday unveiled tools letting Anthropic, Microsoft and OpenAI agents automate tasks inside customer accounts, with usage fees starting in February 2027 after a no-fee grace period. Salesforce has a similar "headless" offering (agents working through its data and connectors rather than its interface). The interesting term is the timing. Once agent actions are billed, finance sees AI as a line item while the benefit still lives in perception surveys. We expect that mismatch to invite budget scrutiny just as expansion plans soften.

Gatekeepers are restricting agent access

JPMorgan said on its Q3 call that it will block external agents, Meta's Muse and Instinct included, from client accounts. Jamie Dimon: "We're not going to allow just any outside agent to come in inside the company because that's just too risky for clients and for JPMorgan." Benedict Evans notes Amazon and other retailers are blocking AI agents, partly over fraud chargebacks and liability. He adds that Amazon's $75bn of trailing ad revenue is invisible to an agent.

Whoever owns the data and the customer sets the terms. That caps how far agents spread in regulated and retail sectors, and it moves value toward identity, permissioning and audit tooling, the unglamorous products that let an incumbent safely say yes to an outside agent.

What would change our view: if the share of buyers tracking AI revenue climbs well above 27% and revenue proof points multiply, we would expect expansion plans to recover. One bank on one earnings call is a lead. It is not yet a trend.

Enterprise AI is still spreading, but 2027 expansion belongs to vendors whose customers can point to revenue rather than hours saved.

What to do

  1. Split each enterprise-AI holding's 2027 ARR plan into productivity-led and revenue-led use cases this quarter. Haircut productivity-led expansion on the assumption that only about one buyer in ten plans substantial increases.

  2. Add a P&L-attribution gate to every enterprise-AI diligence starting now. Require customer references that show revenue gained or contractor and headcount cost removed, plus seat utilization against licenses purchased.

  3. Set a February 2027 checkpoint on Workday's agent metering, tracking price per action, any volume disclosed and customer pushback, as the read-through for SaaS repricing.

The bottom line

Prices tracked one thing: whether an AI asset can show revenue that someone is contractually bound to pay. Floors built on a strategic buyer's goodwill gave way. So did launch discounts and buyers' impressions of time saved, while signed deployments commanded premiums. That breaks the habit of treating capability or a famous suitor as downside protection, and the next markdowns will land on companies whose revenue rests on perception or promotion. Rebuild every AI memo's downside case this week around contracted, attributable revenue, and strike any line that depends on a suitor, a subsidy or a survey.