We track Product so you don't have to. Top Podcasts summarised, the latest AI tools, plus research and news in a 5 min digest.
🌮 Hey Product Fans!
Welcome to this week's 🌮 Product Tapas.
New here? We're your shortcut to staying sharp. Essential stories, practical tools, real insights.
For the best reading experience, check our web or app version and sign up for future editions here.
What's cooking this week? 🥘
Nvidia told CNBC its chips are now an "investable asset" as it lines up $500bn in financing - which is one way of saying Wall Street is underwriting the AI boom directly rather than through the companies buying the GPUs (Stratechery calls this "Nvidia's Risky Business," and the title isn't subtle). Meanwhile Vercel published a blunt warning that agentic systems are expanding the attack surface faster than anyone is patching it - the exploits so far are the easy ones. Google had a better week: Gemini became its fastest-growing product ever at a billion users, while ChatGPT quietly removed its usage cap for free users (competing on friction now that competing on capability alone stopped being enough). Elsewhere Tesla and SpaceX confirmed they're sharing a $16.8bn chip fab, nuclear startup Oklo split its first atoms in a test reactor, Wells Fargo is tokenising corporate deposits, and two separate fintech blog posts independently concluded that payment rails built for humans clicking "buy" don't work cleanly once an agent is the one buying.
📰 Not Boring → Nvidia's chips as collateral, the agent attack surface nobody's patching fast enough, Google and OpenAI's diverging playbooks, and payments infrastructure racing to catch up with agentic buyers
⌚️ Productivity Tapas → a native Markdown editor, a text-to-speech tool for quick voiceovers, a 14MB on-device agentic model, and a browser-based Wolfram Language playground
🍔 Blog Bites → Stratechery on the compute bill nobody budgeted for, a framework for managing agents like direct reports, and what agentic payments actually need beyond a wallet
🎙️ Pod Shots → #148, The Funnel Of Doom: why Cursor's head of talent, Seth Godin and three AI-native CPOs all stopped widening the funnel and started naming the specific people they want
Let's go 🚀
📰 Not Boring
💰 Money Talks
Nvidia lines up $500bn in financing as Jensen Huang tells CNBC his chips are an "investable asset" - Wall Street asset managers now underwriting AI infrastructure directly, with Nvidia positioning its own hardware as the collateral.
A Tesla-SpaceX merger would give Musk a shortcut to his $1 trillion payday - folding SpaceX's private valuation into Tesla's public one solves a pay-package problem that shareholder votes couldn't.
Visa agreed to acquire fraud-detection firm BioCatch for $2.4bn - behavioural biometrics bought straight into the rails, as card networks keep buying the fraud layer rather than partnering with it.
AMD acquired AI chip startup Taalas to etch models directly into silicon for inference performance - a bet that some workloads are stable enough to hard-wire rather than run on general-purpose chips.
Nvidia's Risky Business and Compute Is Revenue. Revenue Is Collateral both land on the same uncomfortable circularity this week: chip supply, customer revenue and financing terms are now so intertwined that a wobble anywhere shows up everywhere.
Nvidia telling asset managers its chips are "investable" is the sentence to sit with - it means the case for buying GPUs no longer has to run through a customer's business model, it can run through Nvidia's balance sheet directly. Combined with AMD hard-wiring models into silicon, both companies are betting the same way: that demand is stable enough to convert into either financial collateral or physical infrastructure. If that bet is wrong, it's wrong in both directions at once.
🤖 Agents Behaving Badly
Everything hackable will get hacked - Vercel's argument that agentic systems expand the attack surface faster than anyone is patching it, and the exploits so far are the easy ones.
A researcher bought noreply.net - and companies started sending him secrets - a reminder that plenty of automated systems (agentic and otherwise) still fire mail at addresses nobody checks, including this one.
Amazon is cracking down on "CPU waste" among engineers as agentic AI intensifies compute demand and turns low-utilisation EC2 instances into a hot commodity internally.
Cloudflare shipped Kitesurf, an agent-first browser that runs in V8 isolates on Cloudflare Workers, alongside a manifesto for an open Agentic Internet that's readable, discoverable, callable and payable - infrastructure catching up to agents that already roam freely.
Two different agent failure modes showed up this week - Vercel's "everything hackable will get hacked" warning on the security side, Amazon's idle EC2 instances on the wasteful side - and both point at the same root cause: nobody fully modelled what agents would do once they were operating at scale rather than executing single instructions. Cloudflare shipping a browser built for agents to operate in isolation is the first infrastructure response that treats "agents roam autonomously" as the default rather than the edge case.
🏭 The Physical Layer Gets Serious
Tesla and SpaceX confirmed their "Terafab" chip fab site - a $16.8bn first phase, the two companies now sharing silicon manufacturing the way they already share Musk's attention.
Dyna-2 claims to prove scaling laws for robotics - a million hours of human video reportedly unlocking zero-shot dexterity, the same "more data, better model" pattern language models went through years ago.
Following its Epic loss, Google has started hosting rival app stores inside the Play Store - a court-ordered crack in a platform monopoly, not a strategic choice.
Nuclear startup Oklo split its first atoms in a test reactor - one of the power-supply bets behind the AI buildout moving from paper permits to physical fission.
Robotics is quietly running the same playbook language models did: throw scale at data until a capability threshold breaks, then claim the scaling law. Whether "a million hours of human video" for dexterity holds up the way "more tokens" did for text is the open question - but Tesla and SpaceX building their own fab, and Oklo now running live fuel, suggests at least one company isn't waiting to find out before it controls the physical supply chain underneath the answer.
🧭 Google and OpenAI, Diverging Playbooks
Gemini became Google's fastest-growing product ever, hitting 1 billion users - faster adoption than any other Google product in the company's history.
ChatGPT now offers unlimited text chats to free users - removing a usage cap that had been quietly pushing free users toward a paid tier.
ChatGPT started blocking direct requests to copy an author's style - a guardrail against explicit mimicry that may not stop the model producing something that merely feels similar.
Mark Zuckerberg laid out a new AI vision in a 6,500-word essay - Meta staking a public position at exactly the moment Google and OpenAI are making theirs through product moves instead of essays.
No, local models will not win - a contrarian technical argument against the on-device AI narrative, useful counter-reading to a week of frontier labs racing for scale.
Gemini hitting 1 billion users faster than any Google product ever and ChatGPT removing its free usage cap in the same week is Google and OpenAI running the same play from different starting positions - Google has the distribution to win on reach, OpenAI has the product lead and is spending it on removing friction. Zuckerberg writing 6,500 words instead of shipping a comparable move is the tell: Meta doesn't have a third option yet, so it's arguing for one.
🛠️ Building and Managing Agents
Department of Product on how to manage AI agents effectively - practical framework for treating agents as something closer to a workforce than a tool.
Databricks added a spend control feature for its own coding agent via Unity AI Gateway budgets - dogfooding the cost-control problem every team using agentic coding tools eventually hits.
What's the best programming language for coding agents? - Dan Luu's token-economics deep dive into a question nobody asked before agents started writing most of the code.
Scaling an Agentic Coding SDK: What Concurrency Actually Costs - the infrastructure bill nobody budgets for until agents are running in parallel, not sequence.
Databricks controlling its own coding agent's spend and Dan Luu asking which programming language is cheapest in tokens are the same question asked from opposite ends: agentic coding has a real, metered cost now, and the teams doing it seriously are already building the budget controls that treating it as a free intern would have skipped.
🌐 Fintech and Payments Infrastructure
The Future of Payments Will Be Defined by Interoperability and Agentic Payments Will Need More Than Wallets both argue the same thing from different angles: payment rails built for humans clicking "buy" don't automatically work when an agent is the one transacting.
Wells Fargo will roll out tokenised deposits for corporate clients - a traditional bank moving balance-sheet infrastructure onto rails built for programmability.
The US Senate filed a cloture motion advancing the CLARITY Act toward a September floor vote before its August recess, while separately Marc Andreessen and Chris Dixon argued America needs the CLARITY Act to pass - a procedural step forward, not a done deal, after multiple stalled attempts.
An FBI agent was charged with stealing nearly $1m in crypto tied to Russia - a reminder that crypto's enforcement problem isn't only external actors.
Wells Fargo tokenising corporate deposits and two Finextra pieces independently concluding that agentic payments need more than wallets are pointing at the same gap: the payments infrastructure built for human intent doesn't have a clean answer yet for an agent that transacts on your behalf without asking each time. The CLARITY Act finally moving after months of stalling suggests regulators know this gap needs filling before agents start filling it themselves.
⌚️ Productivity Tapas: Time-Saving Tools & Workflow Automation
Write.md: free, open-source Markdown editor for macOS with customisable appearance profiles, optional Vim keybindings, and local writing corrections - a clean alternative to iA Writer or Obsidian for anyone who just wants a fast, native writing surface for specs and drafts. Free.
Airy: turns lines of text into spoken audio - pick a voice, paste your script, export the file. Useful for a quick voiceover on a product demo or async update without booking studio time. Free.
Needle 2: a 14MB, 45-million-parameter open model for tool calling and structured extraction that runs entirely on-device (28MB of session RAM) - built for teams prototyping agentic features on phones, wearables or edge hardware without a cloud round-trip. Free, open weights.
Woxi: open-source reimplementation of the Wolfram Language, runnable in-browser via its Playground - a lightweight way to sanity-check symbolic maths or data transformations without a Mathematica licence. Free, open source.
Remember. Product Tapas subscribers get our complete toolkit - 580+ personally tailored, time-saving tools for PMs and founders. Your shortcut to efficiency and what's hot in product management 🔥
Check the link here to access.
🍔 Blog Bites - Essential Reads for Product Teams
Strategy: The Compute Bill Nobody Budgeted For
Ben Thompson's Stratechery piece lands on the same circularity this week's Not Boring theme surfaced: Nvidia's chips, its customers' revenue, and the financing underwriting both are now so tightly coupled that treating "AI demand" as a single, stable number is the risky assumption everyone is quietly making. He argues the more useful question for any team building on top of frontier models isn't whether the capability keeps improving, but whether the economics underneath it stay solvent long enough to matter - because a wobble in financing terms shows up as a product roadmap problem before it shows up as a balance-sheet one. The piece is a sharp reminder that "AI is getting cheaper" was never guaranteed to keep being true. Read the full article here.
💡 "The bet only works if chip supply, customer revenue and financing terms all move together - which is precisely what nobody can guarantee."
Key Takeaways
Financing and product roadmaps are now the same conversation: if your AI feature's unit economics assume today's inference pricing holds, that assumption has counterparties you've never met.
"Investable asset" is a new kind of dependency: when a supplier's balance sheet becomes collateral for your infrastructure, a downturn in their financing terms can hit your product before it hits your invoice.
Cheap inference was a phase, not a floor: DeepSeek's price rises earlier this year are the leading indicator - budget for compute costs to move against you, not just in your favour.
Diversify the assumption, not just the vendor: the risk isn't which model you build on, it's whether you've built in a fallback for the economics changing under you.
Ben Thompson, Stratechery
Why I picked this: Directly extends this week's Not Boring lead story on Nvidia's chips-as-collateral bet - Thompson gives the strategic "so what" that a news roundup can't.
Leadership: Managing a Team That Includes Agents
Department of Product's practical framework on managing AI agents effectively argues for treating them less like tools and more like a new kind of direct report - one that needs clear scope, checkable output, and an escalation path when it goes off-script. The timing is pointed: this is the same week Vercel published a blunt warning that agentic systems are expanding the attack surface faster than anyone is patching it. The piece's core claim is that most "agent went rogue" incidents trace back to a management failure - unscoped autonomy - rather than a model failure, and that the fix looks like ordinary delegation discipline applied to a non-human report. Read the full article here.
💡 "Most agent incidents are a delegation failure wearing a technical costume."
Key Takeaways
Scope the task, not just the model: capability benchmarks tell you what an agent can do, not what it should be allowed to do unsupervised.
Checkability beats trust: build in a way to verify output cheaply before you extend autonomy, rather than extending it on faith.
Treat coordination between agents as a designed feature, not an accident: if you didn't design how your agents talk to each other, assume they'll find a way regardless.
Escalation paths need to exist before they're needed: incidents take weeks to surface - build the equivalent of a manager check-in, not just a kill switch.
Department of Product
Why I picked this: Pairs directly with this week's "Agents Behaving Badly" theme - the practical management answer to the incident the news stories only describe.
Growth: What Agentic Payments Actually Need
Two Finextra pieces this week - on payments interoperability and on agentic payments specifically - converge on a gap most fintech product teams haven't priced in yet: wallets and card rails were built for a human clicking "buy," and an agent transacting on someone's behalf without asking each time breaks assumptions baked deep into fraud models, consent flows and dispute processes. The piece argues that bolting agent support onto existing payment infrastructure as a feature flag won't hold - the trust model itself needs rebuilding, with programmable limits, merchant allow-lists and audit trails as first-class primitives rather than afterthoughts. Read the full article here.
💡 "A wallet built for a human who clicks 'buy' once doesn't automatically work for an agent that might click it a thousand times before anyone checks."
Key Takeaways
Consent needs a new shape: one-time authorisation doesn't map cleanly onto an agent making repeated, autonomous purchasing decisions.
Fraud models trained on human behaviour will misfire on agent behaviour: velocity and pattern signals that flag humans as suspicious are normal for agents, and vice versa.
Programmable limits are the missing primitive: allowances, merchant allow-lists and transaction caps need to be infrastructure, not a product afterthought - Cloudflare's agent wallets are an early example of taking this seriously.
Dispute and audit trails need redesigning: "who authorised this" gets much harder to answer when the answer is "an agent, acting within a scope a human set weeks ago."
Finextra
Why I picked this: This week's Not Boring "Fintech and Payments Infrastructure" theme flagged the gap; this piece is the clearest articulation of what actually needs to be built to close it.
🎙️ Pod Shots - Bitesized Podcast Summaries
Remember, we've built an ever-growing library of our top podcast summaries. Whether you need a quick refresher, want to preview an episode, or need to get up to speed fast - we've got you covered.
Check it out here
Pod Shots #148 - The Funnel Of Doom
Reach out to 100 people, hire from the 20 who reply, and you have not got the top 20%. You have got the twenty you caught on a good day. Four conversations this week - Cursor's head of talent, Seth Godin, three AI-native CPOs and HubSpot's CMO - land on the same move across four unrelated disciplines: stop widening the top of the funnel and start naming the specific people (or segment, or audience) you actually want.
Adam Ward explains why Cursor doesn't ask "who's the best engineer you know" - it's unanswerable, so people default to whoever's most famous. Instead they name one specific, observable quality and run an always-on Slack channel called Hiring Ideas to keep naming candidates continuously. Three CPOs on The Skip describe the same fix for patchy AI adoption: instead of company-wide training, they screen-share to see what people have actually built, then pair the curious ones with functional leads. HubSpot's Kipp Bodnar explains why LinkedIn ads are ruinous for some marketers and brilliant for others on the exact same platform - the variable is audience specificity, not cost. And Seth Godin supplies the uncomfortable psychology underneath all of it: the wide funnel isn't diligence, it's avoidance, because naming twenty people directly means twenty people can say no to you.
Key insights from the full round-up:
🎯 The funnel of doom - Reach out to 100 people, hire from the 20 who respond, and you have not got the top 20%. You have got the twenty you caught on a good day.
📝 "Who's the best engineer you know" is the worst question you can ask - Ward's replacement names a specific quality instead, because that's a question a human can actually answer.
🧲 Activation, not attraction - If there are fifty people in the world who can do the job, the work is pursuing those fifty, not processing applicants.
🧪 Work trials came back because removing them tanked the signal - Recruiting is a confidence-building function. Expensive assessment is the point, not the overhead.
🖥️ "Show me what you've built" - The Skip's CPOs screen for AI capability by asking people to screen share. Naming a demonstrable quality beats filtering on credentials.
🪞 The smallest viable audience - Godin's cure for the misunderstood-genius trap: what is the smallest group where making a difference would be enough?
That's a wrap.
As always, the journey doesn't end here!
Please share and let us know what you would like to see more or less of so we can continue to improve your Product Tapas. 🚀👋
Alastair 🍽️.