Tolan Raised $30M at a $268M Mark. Two Years Ago Google Paid $2.7 Billion for the Same Idea.

Portola raised $30 million in new capital on Oct 19, 2026, led by Union Square Ventures with Khosla Ventures and Homebrew participating. The round doubles the company's total funding to roughly $60 million and comes fifteen months after a $20 million Series A led by Khosla Ventures, on top of a $10 million seed raised in February 2025. Portola did not disclose its valuation. PitchBook estimates it above $268 million, up from a prior mark of $120 million, with the outlet attaching medium confidence to the figure.

The operating numbers behind that mark are real but thin. Tolan has more than 250,000 monthly active users who held over 70 million conversations in the past year, roughly 70% of them women and 40% parents. Annual recurring revenue has more than doubled year over year to what the company calls a solid double-digit-million figure. Net revenue retention after 30 days improved 115% year over year. What is not disclosed is the count of paying users — the single number that would let anyone test whether the growth is conversion or leakage.

The Bank of America Institute reported in March 2026 that roughly 3% of its households paid for AI services, at a median of $20 per month. Read alone, that looks like a hard cap on consumer AI revenue. It is not. The same Institute found that the number of households making AI payments was up 38% versus its 2024 average, and that median spend was up 10.4% year over year. Venture capital firm data published in October 2026 shows the rate itself doubling from 2.1% to 4.5% of U.S. consumers in twelve months.

What actually constrains companion apps is a different number in the same dataset: only 13% of consumers who pay for one AI product pay for a second one. That is a willingness-to-pay-twice ceiling, not a willingness-to-pay ceiling. The spend concentration is extreme — the median AI payer spends about $25 a month, while the top 1% of payers average $903 a month and account for 19.5% of all observed consumer AI spend. Any companion app is therefore competing to become a household's first or second AI subscription, not its ninth.

In August 2024, Alphabet signed a nonexclusive licensing agreement with Character.AI and rehired its two co-founders. Reporting put the consideration at $2.7 billion. In October 2026, The Walt Disney Company agreed to serve as stalking-horse bidder for certain Character.AI intellectual property for $20 million — a perpetual, nonexclusive license. The court papers list the company's assets and liabilities between $50 million and $100 million. The same class of asset imploded from $2.7 billion to $20 million in under two years.

The mechanism matters more than the ratio. Google paid for model capability and talent; Disney paid for character assets and the team that maintained them. The 2026 buyer is not licensing a frontier model — it is buying a distressed library and the engineers who kept it coherent. That distinction tells you what was actually scarce. Tolan's $30 million buys neither a model nor a distribution channel. It buys a character scaffold, a memory corpus, and a first-ever free tier.

The scaffold is the part worth worrying about, because it is precisely what Chapter 11 just put a clearing price on. OpenAI's own customer account of Tolan's architecture describes an in-house science fiction writer and a behavioral researcher authoring and refining each character's scaffold, alongside a memory system that reconstructs the full conversational context on every single turn. Comparable character assets changed hands this month for one one-hundredth of one percent of Tolan's estimated valuation. That is an inference, not a disclosed fact — Portola does not break out how much of its value sits in the scaffold versus the user base — but it is the arithmetic a buyer would run.

The bull case for aggregators is that distribution converts. The cleanest public test ran and failed. On Nov 5, 2025, Snap announced that Perplexity would pay $400 million over one year in cash and equity to put conversational AI search inside Snapchat's chat interface. On May 6, 2026, Snap disclosed the two companies had amicably ended the relationship during the first quarter, and that its sales guidance assumed no contribution from Perplexity. The deal died before a dollar appeared in guidance, having been in testing when the parties said in February they had yet to agree on a broader rollout.

Distribution alone is not the moat the aggregator thesis assumes. Snap had hundreds of millions of daily users and still could not align with a model vendor on terms it would accept. What separates the aggregators that work from the one that failed is not reach — it is the ability to give the model away inside an existing subscription bundle and monetize somewhere else. That is the distinction this article turns on.

Meta launched Muse on Sep 8, 2026 as a personal AI agent with a free tier, a $20 monthly Power plan, and a $100 monthly Maximum plan, distributed through a web app, the iOS and Android stores, WhatsApp, and Meta's glasses. Tolan's legacy pricing ran from $10 to $200 a month with no free tier. Meta's entry planlands squarely inside Tolan's band and arrives with WhatsApp attached, while its free tier removes the paywall entirely.

Tolan's response to that pressure is the free tier this round funds — the company's first. A free tier is a rational defensive move and a poor revenue one. It converts existing paying users into free ones unless new paid conversions arrive faster than the base dilutes, and it pushes Tolan's average realized price toward the $20 to $25 monthly median where60% of all AI payers already sit. Investors underwriting the round at roughly 11x to 27x undisclosed ARR are betting that the funnel widens faster than the price compresses.

Behind both megacaps sits a third layer that captures toll revenue on every one of those conversations regardless of which brand wins. NVIDIA posted $303.0 billion of trailing revenue at a 74.7% gross margin. That margin is the economic residue of inference, and it is the layer Character.AI's $8 million monthly burn was paying out. The companion app competes for a slice of what the chip supplier already owns.

Tolan rebuilt its entire conversational context on every turn — recent message summary, persona card, vector-retrieved memories, tone guidance, live app signals — using OpenAI's GPT-5.1 through the Responses API, which cut speech initiation time by more than 0.7 seconds. Memory recall uses vector lookups returning in under 50 milliseconds, and a nightly job prunes low-value entries. Portola reports that memory recall misses fell 30% and next-day retention rose more than 20% after the GPT-5.1 personas shipped. Those are real product gains. They are also a per-turn inference cost that scales with engagement rather than with revenue.

The scale of the problem is visible in the disclosed numbers rather than modeled. Character.AI lost over $8 million a month at roughly 12 million users. Tolan, at a fraction of that reach, reports ARR more than doubling into the double-digit millions. Both facts can be true simultaneously, and they are: one app reached escape velocity on revenue density while the other reached critical mass on users and died. The disclosed pricing points — a $20 monthly median payer and a $200 top tier inside a category where 7% of payers spend above $100 — explain why revenue density wins. It is the only variable that scales with the inference bill.

The risk to the bullish reading of this round is specific and dated. If Tolan's free tier drives monthly active users up while ARR grows more slowly than the double it posted this year, the implied multiple on a $268 million estimate compresses from double digits toward single digits in a single reporting cycle. Portola would then need the monetization shift it has already begun — meal planning, grocery label reading by video, care coordination for elderly relatives — to convert free usage into paid action. Those features are announced, not shipped, and none of them have disclosed revenue attached.

The risk to the bearish reading is that the free tier works. Consumer AI paid conversion doubled in twelve months; there is no law that companion apps are exempt from that curve. If Tolan's paying density holds as the free tier widens the funnel, the category proves that a narrow, intent-qualified audience converts far better than a general assistant base — in which case the $268 million mark is cheap and the 13% rule is a temporary artifact of a young market.

Stated plainly: the consumer companion trade is real, funded, and working — and the money in it is not where the October funding headlines suggest. Facts first. Tolan raised $30 million, doubled its total capital to roughly $60 million, serves more than 250,000 monthly active users, and doubled ARR year over year. Character.AI entered Chapter 11 with roughly 12 million users, a $20 million license sale, and a burn above $8 million a month. Google paid $2.7 billion for that asset class in 2024; Disney paid $20 million in 2026.

The inference is that this round prices a companion app as a distribution asset while the market is discovering companion apps are compute liabilities with a thin subscription attached. The evidence is the repricing itself. A category where the clearing price for the leader's intellectual property fell 99.3% in twenty-four months does not support a $268 million mark on a smaller competitor's user base, unless the smaller competitor is materially better at something other than reach — and the disclosed numbers say it is better at paying density, which the free tier is about to test.