Home
Jun 21, 2026
View All

AI Teams, AI Spending Data, and the Talent Shuffle at the Top

·1 underrepresented voice

The Short Version#

Three things worth your attention today: Lenny's interview with Fiona Fung surfaces what it actually looks like to run an AI-native engineering team at Anthropic, Stripe published real consumer spending data on AI subscriptions from 250 million customers, and Google's AI talent is walking out the door fast enough that Dare Obasanjo noticed.

Lenny Rachitsky / Fiona Fung - What Running an AI-Native Team Actually Looks Like#

Source: https://www.lennysnewsletter.com/p/building-the-most-ai-pilled-engineering Credibility: High (Lenny's Podcast, first-party interview with the manager of Claude Code and Cowork teams at Anthropic)

What happened: Fiona Fung, who manages the Claude Code and Cowork teams at Anthropic, sat down with Lenny to talk through what it looks like to run engineering when roles are blurring and agents are handling meaningful parts of the work. This is one of the few first-person accounts from inside one of the companies that is both building AI tools and using them to build more AI. It's not theoretical.

Key patterns:

  • "Roles are blurring" is the core challenge she names. When agents do more of the implementation work, the lines between PM, engineer, and product designer start to collapse. What does coordination look like when the person writing the spec and the person (or agent) doing the work are sometimes the same?
  • Culture maintenance becomes harder when the team is growing, agents are everywhere, and what "doing the work" means has changed. She's managing this in real time at a company that is also the source of the tools causing the change.
  • The framing of "AI-pilled" is deliberate. This isn't a team that uses AI reluctantly. It's a team that built their entire workflow around it and is now figuring out what breaks at scale.

Why it matters for PMs: This is what product teams at frontier AI companies actually look like from the inside. If your organization is trying to figure out how to structure teams as AI takes on more execution, Fiona's experience is a closer signal than any consultant's framework. The specific question of "how do you maintain culture when roles are blurring" is one that most product leaders haven't answered yet, and she's living it.

Critical questions:

  • What's the ratio of human to agent work on the Claude Code team, and how does that change what senior ICs actually do?
  • When agents handle implementation, does PM scope expand (more strategy, more context-setting) or contract (less needed overall)?
  • How do you do performance reviews when output is increasingly hard to attribute?
  • Is "AI-pilled" a sustainable culture posture or a phase that gives way to something more structured?

Action you could take today: Listen to the episode and share the framing of "roles blurring" with your engineering leads. Ask them where they see it happening on your team already. The conversation will surface real friction before it becomes a structural problem.

Stripe - What 250 Million Customers Are Actually Spending on AI#

Source: https://stripe.com/blog/industry Credibility: High (first-party data post from Stripe's Link consumer product team, June 18, 2026)

What happened: Dan Hill, PM on Stripe's Link Consumer Product, published an analysis of AI spending patterns across 250 million customers on the Stripe network. This is the kind of data that almost never gets published because companies that have it usually consider it proprietary. Stripe has it because Link sits on top of a massive volume of subscription and purchase transactions, and AI subscriptions are a growing share of that.

Key details:

  • The dataset is 250 million Link customers, which is a real sample, not a survey.
  • The post analyzes AI spending specifically, meaning subscription patterns for AI products (ChatGPT, Claude, Perplexity, etc.) across a broad consumer base.
  • Stripe can see not just whether someone subscribes, but spending patterns over time, which means they can speak to retention and expansion, not just acquisition.

Why it matters for PMs: Consumer willingness to pay for AI is one of the open questions PMs building AI products are still trying to answer. Stripe just published one of the only real datasets on this question. It's worth reading for two reasons: first, for the actual data on what categories of AI products consumers are paying for; second, as a template for how to frame consumer spending analysis when pitching AI investments internally.

Critical questions:

  • Does the data distinguish between active subscribers and lapsed ones? Acquisition and retention look very different.
  • Which AI product categories show the strongest spending retention, and which show early drop-off?
  • Is per-customer AI spending growing over time (expansion) or concentrated in one-time trial-then-cancel patterns?
  • How does Stripe's customer base (likely skews toward online-purchase-literate consumers) compare to the general population?

Action you could take today: Read the Stripe post and pull out the specific categories showing growth. Then check whether your product or your company's AI features fall into a high-retention or high-churn spending category based on what they describe.

Dare Obasanjo - Google's AI Talent Is Walking#

Source: https://mas.to/@carnage4life Credibility: Medium (Dare is a credible industry observer with a product and engineering background; these are commentary posts, not primary sources)

What happened: Dare flagged what's become a notable pattern this week: Google lost its Gemini co-lead (Noam Shazeer) to OpenAI, and separately the AlphaFold project lead to Anthropic. He also noted Zachary Lipton's departure from Google DeepMind. Three significant departures from Google's AI organization in close proximity.

Key details:

  • Noam Shazeer's move to OpenAI is the highest-profile departure. Shazeer is one of the original Transformer paper authors and a foundational figure in modern LLM development. His leaving Google (for the second time) is not a routine attrition event.
  • The AlphaFold project lead going to Anthropic is a different signal: that's someone moving from a non-language-model research context to a frontier model company, suggesting Anthropic is actively recruiting from Google's broader AI research base.
  • Zachary Lipton's DeepMind departure adds to the pattern.

Why it matters for PMs: Talent flows in AI are a leading indicator of capability concentration. When foundational researchers move to OpenAI and Anthropic, those companies get access to knowledge that isn't yet in any paper or product. For PMs thinking about model selection, build vs. buy, or vendor risk: the gap between frontier model companies and everyone else may widen if this pattern continues. Dare's separate observation about OpenAI's operating losses scaling with revenue (losses growing as revenue grows) is a useful counterweight — it raises real questions about the sustainability of the frontier model race that affects everyone building on top of it.

Critical questions:

  • Does losing researchers of this caliber affect Google's ability to close the gap with OpenAI and Anthropic at the model level, or is Google's advantage structural (compute, data, distribution)?
  • If the frontier model concentration continues, what's the product risk for teams building on Google's Gemini APIs?
  • OpenAI's growing operating losses: at what point does that become a product risk for companies deeply integrated with their APIs?

Action you could take today: If your product relies heavily on a single model provider's API, spend 30 minutes stress-testing your vendor concentration risk. What would it cost to swap providers? What's the migration path? Talent shifts like this are a reminder that the frontier model landscape is still unstable.

Quick Hits#

  • Lenny Rachitsky / Community Wisdom: June 20 community roundup includes a thread on sharing Claude Code context across a team — a very practical problem if you're running an AI-assisted engineering workflow. (2026-06-20): https://www.lennysnewsletter.com/p/community-wisdom-fractional-cpo-compensation

  • Stripe Projects: Added new agent integrations, more provider support, and custom developer controls (June 11). Separate from the spending analysis post — this is a product change to their agent-infrastructure tooling for developers. (2026-06-11): https://stripe.com/blog/product

  • ElevenLabs: The Government of Poland took a stake in ElevenLabs through Vinci/BGK Group, joining a16z, Sequoia, and ICONIQ. Government-as-investor in AI voice infrastructure is a new pattern worth watching for anyone thinking about public sector AI adoption. (2026-06-18): https://elevenlabs.io/blog/poland-invests-in-elevenlabs

  • Ravi Mehta: "From 10x people to 10x teams" — analysis of how AI has accelerated individual output but created new coordination bottlenecks at the team level. Connects directly to the Fiona Fung interview themes above. (2026-06-02): https://blog.ravi-mehta.com/p/from-10x-people-to-10x-teams

  • Simon Willison: Datasette Apps — a new feature that lets you host custom HTML applications inside Datasette. Small-team tool builder shipping product; useful pattern for PMs thinking about lightweight data tooling. (2026-06-18): https://simonwillison.net/2026/Jun/18/datasette-apps/#atom-everything

The Thread#

The frontier AI talent market is making vendor concentration risk more real. This week's Noam Shazeer move and the AlphaFold departure from Google, combined with ongoing questions about OpenAI's operating loss trajectory, point to an unstable competitive structure at the model layer. At the same time, Fiona Fung's interview shows what actually building on that unstable layer looks like from the inside. The teams closest to the frontier are managing roles, culture, and workflow norms that nobody has written the playbook for yet — and the rest of us are six to twelve months behind them in figuring out the same problems.

Sit With This#

Fiona Fung described running a team where "roles are blurring" as agents take on more of the implementation work. At Anthropic, the people building Claude Code are also the primary users of Claude Code — which means they're discovering the workflow breakdowns in real time.

For your team: Where is role ambiguity already appearing as AI takes on more execution work? And who on your team is responsible for noticing it before it becomes a coordination failure?