Circle’s valuation is getting nutty

Jun 20, 2025

Welcome to Cautious Optimism, a newsletter on tech, business, and power.

Happy Friday, friends! Slowly, the world is moving towards phone-free schooling (focus), old-fashioned hand-written exams (anti-cheating), and less social media access for teens (mental health). The latest comes from Australia, which is making some progress towards banning kids under the age of sixteen from social media. A decade ago, hackles would have been high that kids’ speech rights were being infringed. Today? I think enough people have seen the impact that unfettered access to electronic and digital entertainment has on kids.

The sticking point is how to verify kids’ ages. Apple and Google don’t want to do it at the app-store level, but that doesn’t leave many other gates to keep. In Australia, getting age-verification right is taking time. — Alex

📈 Trending Up: Jet HR … datacenter buildout … AI cost control? … Meta’s AI spending … Poolside … legal threats … the Israel-Iran conflict … Vietnamese GDP … our agentic decade … people not liking Sam Altman …
📉 Trending Down: The Russian economy … Scale AI’s pre-Meta trajectory? … US-India relations … bitcoin transaction volume …
📒 Round of the Day: Axios’ Ryan Lawler wrote up Payabli’s $28 million Series B, noting that the embedded payments startup’s round was led by Fika Ventures and QED. Payabli’s software allows software companies to offer payments for their customers, will all complexity abstracted away via an API. Why is it our Round of the Day? Because it appears the fintech boomlet is still cresting.

Circle’s valuation is getting nutty

The 2025 IPO class is doing well. CoreWeave is worth $170 per share up from its IPO price of $40 per share. eToro and Chime are up more modest percentages, but are still worth more today than they were when they listed.

Meanwhile, Circle is following in CoreWeave’s footsteps, worth $227.89 in pre-market trading — up 13.8% before the bell today after gaining 33.8% yesterday — after listing at $31 per share.

The company has received good news of late, with the GENIUS Act passing the Senate and heading to the House. But we’re talking about a company that generated $578.6 million worth of gross revenue in the first quarter that now has to support a $48.4 billion valuation. After Circle pays out Coinbase for the USDC it holds on its platform — the two companies split USDC economics — the ratios look even more wonky.

But don’t think that we’re here to scold. The opposite. Market excitement for tech IPOs generally, and more specifically in the crypto-banking space, is great news for startups building in the same area. Talk about a positive comparable!

Tesla’s robotaxi business to start modestly

Currently pipped for a June 22nd launch, Tesla’s much-discussed robotaxi business will start life with around ten cars, the FT reports, and will be constrained by a geo-fence “to avoid [Austin’s] most challenging intersections and come with backup teleoperators poised to intervene if problems occur.”

For a new robotaxi service, the restrictions seem reasonable and prudent. Waymo also takes a very cautious approach to new markets. The Alphabet subsidiary has announced a handful of new cities it intends to reach in 2026, for example, giving it ample time to prepare for the new terrain.

Telsa fans are receiving invites to the service

I don’t think the precise number of self-driving taxis that Tesla starts with matters; nor does the initial driving area matter much. Instead, I care lots about how quickly the company can scale from a conservative starting point; in self-driving, you have to go slow to go fast. Acceleration, then, is the name of the game.

Waymo announced recently that it has New York City on its list of future markets. Along with planned expansion into Washington D.C. and we’re starting to see self-driving cars approach cities with real winters. Encouraging.

Cloudflare plans to shake up AI-Publisher economics

The working title for this section was Yeah, we need to tell the AI companies to fuck off, but in a fit of maturity, I changed it.

New data from Cloudflare backs up what we learned from TollBit: AI model companies are scraping more, and sending fewer visitors back to source material over time. TollBit’s data showed that AI companies are now scraping more for RAG uses than model training needs, indicating that AI users are asking questions that require online searching to provide answers. The bad news for websites, however, is that as AI traffic soars, the number of people sent back to sources is de minimis, per the IP-AI marketplace.

Cloudflare agrees. Axios reports that a decade ago, Google would crawl two pages “for every visitor it sent a publisher,” citing the internet infra company’s CEO Matthew Prince. In the AI era, the ratios are very, very different. Google reached a 6:1 scraping to vistor-sending ratio six months ago; today it’s eighteen to one. OpenAI went from 250:1 to 1,500:1, and Anthropic offers publishers even worse economics.

In simple terms:

In the Google search era, Google would index your website and send you traffic. The exchange was simple: Websites ate the costs of being scraped, Google made billions off search, and in return it sent a winsome, steady stream of humans to source websites.
In the AI search era, the economics have shifted dramatically. Now websites have to handle more total scrapes (which isn’t free) and in return receive all but no humans coming to their website. This means that the value derived from content work has gone from ‘mostly acceptable because how else would search work’ to ‘our data is being stolen for off-site monetization and we get nothing back for covering scraping costs.’

Most pernicious — and we’ll be brief here since I know we’re on well-trod ground — AI companies are treating human-derived RAG queries as non-machine. In other words, for a great many AI searches, the provider company has exempted themselves from following website rules meant to keep machines off their IP. The situation is an aggressive wiggle around good taste, and the rules.

Perhaps if AI companies had offered better terms early on they could have kept their access to the open web. But they have not, mostly, and thus we’re about to see the world change. The same Axios story reports that Cloudflare is building a “a new tool that will stop content scraping,” with Prince saying that he goes “to war every single day with the Chinese government, the Russian government, the Iranians, the North Koreans, probably Americans, the Israelis, all of them who are trying to hack into our customer sites,” and thus is more than capable to “stop some nerd with a C-corporation in Palo Alto.”

Even more, Cloudflare claims that major publishers are interested in its upcoming tool.

In a single stroke, Cloudflare could — if it can live up to its promises, and frankly I think that it can — section off much of the Internet from AI companies that want to lever online human knowledge sans payment. We could therefore see an arm race between Cloudflare and other companies offering scraping blockers, and AI companies that don’t want to pay for their raw materials.

Or the market could shift towards TollBit, Human Native, Created by Humans and their competitors in building IP licensing marketplaces for AI use cases. But I doubt it; free is one hell of a drug to give up on.

Some folks don’t mind the scraping, of course. Brands that don’t mind scraping costs and want their products to exist in an AI search context, for example. That’s probably why Profound — a startup that helps others with AI search visibility, the modern equivalent of SEO — just raised $20 million.

New TWiST500 companies

Yesterday we promised to share the next set of companies joining the TWiST500, the list of the top ~1% of startups in terms of their likelyhood of generating incredible returns for their backers; it’s a good proxy for total market disruption the startups may bring as they scale.

I still need to do a final cull, but the next cohort of additions include: Mansa, PostHog, Harbinger, Applied Intuition, Mercor, Maven, Turing, Labelbox, Handshake, SuperAnnotate, Tennr, Faces.App, SandboxAQ, Chorus.sh, Mechanize, TerraPower, and Teamworks.