AI Startup Funding Concentrates as Stripe, Thrive Bet on Scale

Three pieces, one theme

This week's venture newsletter cycle bundled together three pieces that look unrelated at first: Stripe's economists laying out a research agenda for AI, a close look at how Thrive Capital builds its brand, and a warning from Angular Ventures that AI startups are converging on the same product.3 Read side by side, they point to one shift in AI startup funding. The technical building blocks of AI products are quickly becoming commodities. The money is going to whoever owns scale, proprietary data, or relationships that keep paying off.

The macro numbers make the case on their own. US venture deal value reached $515.8 billion through Q3 2026, about 44% above the full-year record set in 2021, and AI took a record 82.7% of it.34 Corporate investors account for a record 87.9% of US AI deal value this year, even as they take part in a shrinking share of deals.35 Crunchbase measures differently but sees the same pattern: a record 27 companies raised rounds of $1 billion or more in Q3, and together they took about a third of all global venture money.

The two trackers disagree on how dominant AI is. Crunchbase puts AI at 64% of global funding in Q3, down from the previous two quarters. PitchBook's US year-to-date figure is far higher.34 The gap mostly comes from geography and time window, not a real conflict. Both point the same way: capital keeps concentrating.

Everyone is building the same thing

Angular Ventures argues that many AI-native startups now run on the same foundation of agent loops, tool connectors, memory layers and sandboxed runtimes. That makes both their products and their pitch decks hard to tell apart.3 Its comparison is to the web after 2005, when user authentication and databases became standard building blocks rather than reasons to choose one product over another. On this view, durable advantage comes from deep industry expertise, customer relationships, and vertical software built into complex enterprise workflows, not from custom prompt harnesses or clever token tricks.30

The argument extends a theme the firm raised last December. It warned then that startups had fallen into "consensus incrementalism," with whole categories funded heavily despite having little to set them apart.31 The engineering world is showing the same convergence from below. Developers now describe the agent stack as three distinct markets (loop frameworks, durable runtimes and sandbox harnesses), each with several interchangeable vendors.25 Most mid-sized tech companies are reportedly building their own internal harnesses for agent sandboxes and permissions.30 When every buyer can build the scaffolding in-house, selling the scaffolding gets harder.

Mitch Joel made a similar point about the frontier labs earlier this year. In his view, OpenAI, Anthropic, Google and Meta are all heading toward the same general-purpose intelligence layer, so the real question is less about whose model is best and more about who can build a lasting business around it.1

The OpenRouter exception that proves the rule

Stripe's purchase of OpenRouter seems to cut against the commodity thesis. OpenRouter is plumbing: a gateway that sends requests to more than 400 models from over 80 providers.14 Stripe did not disclose the price, but the New York Times reported $7.5 billion, compared with a $1.3 billion valuation in May. The founders alone reportedly receive $1.5 billion, and Stripe reportedly outbid suitors including Databricks.16 Reports of the price have varied. The Wall Street Journal described talks at around $10 billion in July.18 Bloomberg later reported a deal above $7 billion.20

Why would a router be worth that much if routing is easy to copy? The likely answer is that Stripe did not buy routing code. It bought a network and the data that came with it. OpenRouter handles more than 10 trillion tokens a day for about 10 million developers and companies, and charges a 5.5% fee on credit purchases.19 Menlo Ventures, which led its Series A, says the company's main product is not routing but being the best gateway to many models. Its moat is scale: a large dataset of prompts, model choices and outcomes, plus enough clout that new models often launch on OpenRouter first.17 Menlo also notes that more than ten companies, including Ramp and Cursor, recently launched their own routers.17 The feature is a commodity. The network built around it is not.

That reading fits Angular's thesis rather than contradicting it. A PitchBook analyst described the deal as Stripe's attempt to put itself in the middle of AI-era capital flows, giving it some leverage over frontier labs, hyperscalers and neoclouds.16 Not everyone sees that as purely good news. One startup lawyer warned that companies that chose OpenRouter to avoid depending on a single provider now rely on one company for model access, billing and routing decisions, and should renegotiate their contracts accordingly.21

Stripe's data play

Stripe's research agenda shows the strategic logic most clearly. Its economists say Stripe serves 88% of the Forbes AI 50 and plan to use payments, business-formation and survey data to study AI's effects on productivity, labor markets, firm structure and agent-driven commerce.2 The agenda openly assumes future access to OpenRouter data and notes that the deal had not yet closed as of October 8.2 That is several weeks past the early-September timeline implied when the deal was announced in August.16

Several of the questions Stripe asks matter directly for valuations. It notes that falling intelligence prices could lift productivity for AI users while cutting revenue for AI providers. It also asks whether AI will produce more, smaller firms or strengthen returns to scale.2 The pricing pressure is already visible. Harvey's gross margins reportedly fell to -50% in June because usage costs climbed under flat per-seat pricing, which pushed it toward a cheaper fine-tuned open-weight model. Legora, meanwhile, is moving toward consumption-based pricing and model routing.3 That is the demand behind OpenRouter's price: when inference costs can sink a startup's margins, whoever controls cost optimization has pricing power.

How Thrive does it

Thrive Capital may be the clearest example of concentration paying off. Laurie Owen's analysis argues that Thrive does almost no traditional marketing. Instead, it relies on founder referrals, bold deals and carefully designed physical touchpoints, and it recently hired a storyteller for Thrive Holdings.13 Owen also notes that Josh Kushner's family wealth and connections are advantages that are hard to copy.13

The track record explains the confidence. Thrive first backed OpenAI in January 2023 at a $29 billion valuation and closed a fund of more than $10 billion in February that was heavily oversubscribed.12 It makes only about a dozen new investments a year and avoids backing rival companies, a stance Kushner calls "ride or die."12 It first invested in Cursor at a $400 million valuation, and 14 months later SpaceX acquired Cursor for $60 billion.8 Assets under management have reportedly passed $65 billion, up from about $50 billion at the February close.812 Thrive Holdings has raised more than $2 billion to buy accounting and IT-services firms and rebuild them around AI, and it was valued at $12 billion in August.8

Thrive Holdings is worth a closer look in light of Angular's argument. If AI scaffolding is a commodity, one way to capture value is to stop selling software and buy the customers instead. In that model, the firm owns the workflow, the proprietary data and the margin gains directly. Thrive is betting the most durable AI value sits inside existing businesses rather than in another tool sold to them.

The liquidity problem

The less flattering side of this boom is in the exit data. There are now 992 active unicorns worth a combined $5.7 trillion, and PitchBook says there is no clear path to liquidity for most of them. Airtable and Miro both sold at discounts of nearly 90% to their prior valuations.34 Fund distributions are running at 7.9% of net asset value, compared with a long-run average of 14.5%.37 Even with SpaceX's record IPO, only 18 companies went public in Q3.34

That gap explains why fundraising keeps consolidating around firms like Thrive and Andreessen Horowitz.34 It also explains why acquirers with strategic reasons are setting prices for the wider market. Nvidia's $12.9 billion bid for Hugging Face and AMD's $8.2 billion deal for World Labs follow the same logic as Stripe's purchase.

The verdict

Taken together, the coverage supports a clear conclusion: in AI, being undifferentiated is costly. Startups that resell the standard agent stack face squeezed margins, look-alike pitches and an exit market that can value them at a fraction of their last round. The winners are those that own something hard to copy, whether a network of developers, a body of transaction data, or a run of concentrated bets that founders vouch for. The OpenRouter deal is not proof that AI plumbing is valuable in general. It shows that plumbing is valuable only once it has become a network.