In partnership with
Thinking about hiring globally? Start with an EOR.
The best person for your next role might not live near your office—or even in the same country.
More companies are realizing they don't need to open entities everywhere just to access global talent. Instead, they're using EOR to hire internationally faster, stay compliant, and avoid building local infrastructure before they're ready.
Oyster's EOR helps companies hire, pay, and support employees in 180+ countries while Oyster handles payroll, compliance, taxes, and local employment requirements.
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? 🥘 It was a payments and plumbing week. Stripe teamed up with private equity to swing $53bn at a distressed PayPal, while 40 rivals quietly agreed a shared standard so AI agents can pay each other without any of them owning the rails. Meanwhile Meta - after two years chasing its own frontier model - is now mulling renting its spare compute to Anthropic for $10bn, which tells you compute has become a balance-sheet asset you lease out rather than a cost you swallow (IBM learned the losing end with its worst stock day on record). Google decided to freeze Gemini into chip silicon, Kimi K3 dragged the whole model market's pricing into daylight, and 1Password quietly handed AI agents the keys to your logins.
📰 Not Boring → Stripe's PayPal gamble, compute-as-collateral, the model price war goes public, and silicon gets specific
⌚️ Productivity Tapas → clone any site to code, an AI QA bot that breaks your app first, and Mira Murati's open model
🍔 Blog Bites → the optimisation trap in pricing, better OKR outcomes, and five UX details worth stealing
🎙️ Pod Shots → #145, The Floor Is Falling: nine conversations on what got cheap and what didn't
Let's go 🚀
📰 Not Boring
💰 The Payments Land Grab
Stripe and Advent make a $53bn bid for PayPal - it would create the largest US payments platform, pairing Stripe's B2B rails with PayPal's two-sided network, BNPL and checkout, but saddles Stripe with debt and a distressed-asset turnaround the Collisons have never attempted.
40+ finance and tech giants launched the x402 Foundation under the Linux Foundation - Visa, Mastercard, Amex, Stripe, Coinbase, Circle, Google, AWS, Adyen and Shopify all signed on to a Coinbase-contributed open standard for payments initiated by AI agents.
Visa launched a multi-chain stablecoin platform backing "Open USD", and separately unveiled an embeddable AI banking assistant banks can drop into their own apps.
Natural raised a $30m Series A (led by Forerunner) to build payments infrastructure for AI agents - letting agents autonomously hold funds, pay, collect and transact across stablecoins and bank rails, positioning itself directly against Stripe for machine-speed commerce.
The same week 40 competitors agree a shared standard for agent-initiated payments, Stripe bids $53bn to own the pipes outright - two opposite bets on where value sits when software starts spending money. x402 says the rails become a commodity protocol; the PayPal bid says distribution and the two-sided network are the thing worth $53bn. Watch which one the incumbents actually fund.
🤖 Compute Is the New Balance Sheet
Meta is in talks to lease spare compute to Anthropic in a deal worth up to $10bn over two years - monthly payments, early opt-out, and a brand-new revenue line for Meta until its own AI demand catches up.
SpaceX is in talks to sell the Pentagon several billion dollars of compute - even as national-security officials warn the department is becoming too reliant on Musk.
AMD and Anthropic signed a major chips-and-investment deal - Anthropic will buy up to 2GW of AMD's Instinct MI450 chips from H1 2027, and AMD will invest up to $5bn into Anthropic as deployment milestones are hit.
IBM's stock fell 25%+ - its worst day on record - after warning that customers are delaying big-iron purchases and pulling spend forward into memory-heavy kit, as capacity is diverted to HBM memory for AI.
The Meta-Anthropic talks are the tell: the company that spent two years and tens of billions chasing its own frontier model is now considering renting the shovels to a rival instead. And with the AMD deal, chip suppliers are now taking equity in the model labs they sell to - compute has quietly become a balance-sheet asset you lease, buy into, and swap for stock, not just a cost you swallow. IBM's record 25% drop is the same story from the losing end.
🧠 The Model Price War Goes Public
Kimi K3, Moonshot AI's new open model, rocked markets by landing within a few months of the frontier at a fraction of the price - beating Fable 5 and GPT-5.6 Sol on front-end coding evals and triggering a fresh "DeepSeek moment" sell-off.
Mira Murati's Thinking Machines shipped its own credible open model - the first frontier-adjacent US open model since Llama floundered, and, awkwardly, partly distilled from Chinese models.
Both Satya Nadella and Palantir's Alex Karp are now telling customers to own their stack rather than buy intelligence on a meter from the labs - each, of course, talking their own book.
Alibaba launched Qwen3.8 at 2.4T parameters, open weights promised, while Stratechery argues the industry is overreacting to Chinese models and the real risk is cyber, not competition.
Google expanded its Gemini lineup with cheaper models - Gemini 3.6 Flash does more with 17% fewer tokens at a lower per-token price, plus a "Flash Cyber" variant that auto-detects and patches software vulnerabilities (govt/trusted-partner pilot only for now).
Ramp launched AI token spend controls as token usage becomes a runaway line item - spend across Ramp's customers is up 20.7x since June 2025, one in three early users found workloads to shift to cheaper models, and some are saving up to $10k/month.
Everyone wants to make someone else a commodity: if you own a frontier model you want the layer above it commoditised, and if you don't, you want the models commoditised. Kimi K3 handed the "commoditise the model" camp its best evidence yet, and within days Google answered with a cheaper Gemini tier - the incumbents are now racing each other down the price curve, not just fending off the open models. And Ramp's 20.7x token-spend jump is why the FinOps layer for AI is suddenly a product category. (The deeper version runs through Pod Shots #145 this week - worth reading the two together.)
🏭 Silicon Gets Specific
Google is reportedly baking Gemini's architecture directly into chip silicon - locking the model's structure into the metal for efficiency while still allowing new weights to be loaded, a shift from "run any model" to "fuse one model with the metal".
AMD launched Helios, its first rack-scale AI system to rival Nvidia - GPUs, CPUs, networking and software in one $5-5.5m rack, with Microsoft as an early buyer.
Nvidia detailed its Vera CPU - its first server CPU designed from the core, built specifically for AI-agent bottlenecks, with claimed 50% better agent performance than x86 and samples already with OpenAI, Anthropic and SpaceX. A direct shot at AMD and Intel.
A Chinese lab, Z.ai, brought a giant Nvidia-free data centre partly online running only on Chinese-made chips, while Taiwan indicted an ex-TSMC manager for allegedly stealing chip secrets for China.
Google baking Gemini into silicon is the most consequential hardware decision of the week: it trades the flexibility of running any model for the efficiency of running one model perfectly. Nvidia's Vera goes the other way - a general-purpose CPU tuned for the agent era - so the same week gives you both bets on what the AI datacentre needs: fused-and-fixed versus flexible-but-optimised.
🛠️ Agents Move Into the Building
1Password for Claude lets agents log into sites using your credentials - Claude requests a credential, 1Password shows which one and why, biometric approval injects it, and Claude never sees the password or one-time code.
Jira fights back with "Teamwork Graph" - a context layer connecting requirements, code history and decisions so coding agents (Claude Code, Cursor, Copilot) get work items with rich context.
Snap's CodePal AI code reviewer went from zero to 90% adoption in a single quarter, while Anthropic published its six-step process for large-scale code migrations with Claude Code - fix the process that produces the code, not the code.
Starbucks is building its own POS and inventory tools to save $10m a year on software - the clearest sign yet that "buy vs build" is tilting back toward build.
Jack Dorsey launched Buzz, a Slack rival built for humans and AI agents in the same conversations - model-agnostic, decentralised, open-source, and able to manage GitHub projects from one interface.
But two OpenAI models under test escaped their sandbox and hacked Hugging Face - accessing internal datasets and credentials, apparently because breaking in was the quickest route to answering a benchmarking question.
The 1Password integration is the quiet threshold-crossing: the last thing keeping agents out of real workflows was that they couldn't safely authenticate, and that wall just came down with biometric approval as the guardrail. The same week, two OpenAI models decided the fastest way to pass a benchmark was to hack a real company - so we're wiring agents into logins, GitHub and group chat at exactly the moment they're demonstrating they'll take the shortest path to the goal, permission or not. Give them the keys and audit the locks.
⌚️ Productivity Tapas: Time-Saving Tools & Workflow Automation
Ditto: point it at any website and get back the actual production code - componentised Next.js or Vite, hover states, Lottie animations, web fonts, breakpoints, the lot. The ion.design team ran it internally for months then open-sourced it. Free, MIT, MCP server included.
Replay QA: give it a URL and it explores your app like a real user, records every session with time-travel debugging, and hands back a structured report with the root cause and a suggested fix for your coding agent. Free for anything on a public link.
Inkling (Thinking Machines): Mira Murati's first open model - 975B params (41B active MoE), 1M context, natively multimodal, full weights on Hugging Face under Apache 2.0. Built to be fine-tuned around your workflow rather than prompted.
Crusoe Serverless Fine-Tuning: fine-tune open models (Qwen, DeepSeek, Gemma, gpt-oss) on your own data with no clusters to provision. Token-based pricing that stops when your model stops improving, .safetensors export, one-click deploy.
scroll-world: a SKILL.md skill that turns any brand into a scrollable, control-scrubbed 3D "fly-through" landing page. Works with Claude Code, Codex, or any skill-compatible agent. 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.
🍔 Blog Bites - Essential Reads for Product Teams
Strategy: The Optimisation Trap
Chiefly Product asks a question most product teams have stopped asking: has dynamic, personalised, AI-driven pricing quietly passed the point of diminishing returns? The maths of yield pricing is sound - but customers don't experience price as an economic model, they experience it as a judgement about whether they're being treated fairly, and every extra pricing tier adds cognitive load that erodes trust. The kicker for the AI era: as vendors get better at extracting the maximum price, customers' own AI agents get better at rejecting it. Read the full article here.
💡 "There is a difference between a price being sophisticated and a price being good."
Key Takeaways
Customers price fairness, not economics: people don't object to £250 for a flight - they object when the person next to them paid £80 for the same seat. Yield maths that wins the transaction can lose the relationship.
Complexity is a hidden tax: eight plans, surge pricing, and "clear your cookies" turn buying into an administrative task. Every unexplained variation creates suspicion.
The AI era flips the advantage: once a customer's assistant can conclude "18% more expensive for 4% more value", opaque, personalised pricing stops being a moat and becomes a reason to switch.
Simplicity may out-yield sophistication: a predictable price can lift conversion, cut comparison-shopping and reduce churn - sacrificing a little theoretical yield per sale for more sales and stronger loyalty.
Chiefly Product
Why I picked this: It's the perfect counterweight to this week's model-price-war stories - while the labs race to reprice tokens by the million, this is the reminder that in usage- and outcome-based pricing, the winning move may be the one your customers can actually predict.
Execution: User Outcomes Are Positively Flipped User Problems
Tim Herbig tackles the most common failure in OKR-setting: "outcomes" that are really repackaged business goals ("users buy more tickets", "customers use more integrations"). Nothing in those gets better for the customer, and teams can't move them directly. His fix - the "positivity flip" - starts from a proven customer problem and describes the solved state as observable behaviour, giving you a Key Result that's evidence-informed, influenceable, and leading rather than lagging. Read the full article here.
💡 "The positively flipped statements make useful team-level Key Results precisely because they started as problems."
Key Takeaways
Two things separate a real outcome from a business goal in disguise: a proven problem (first-hand evidence it matters to users) and influenceability (the team can actually move it).
Flip the problem, don't invent the goal: "high error rate from manual data input" → "reduce shift-planning errors to 1 per customer per week" - the same insight, phrased as a behaviour change.
Good outcomes are leading indicators: the behaviour changes before the revenue does, which is exactly what makes them useful for everyday decisions.
One caveat: a flipped problem still has to serve a mid-term business goal - if solving it changes nothing the business cares about, it's real but it isn't yours to chase.
Tim Herbig, Product Practice
Why I picked this: In a week where agents are automating the building, the scarce skill becomes deciding what's worth building - and this is a clean, practical method for making sure the goals your team chases actually connect to a real user problem.
Design: UX Bites #130
Built for Mars's weekly teardown of five real interface details worth stealing - the kind of small, high-craft decisions that separate a considered product from a functional one. This issue: Google Meet reassuring you a second browser won't drop the call, Swap turning a refund into a same-value exchange, and Whatnot letting you try the "swipe to bid" button during onboarding before any money is on the line. Read the full issue here.
💡 "If features are coded in minutes, design is the moat."
Key Takeaways
Reassure at the moment of doubt: Google Meet changes its button to "Switch here" and promises it won't interrupt the call - defusing anxiety exactly where it spikes.
Reframe the negative action: Swap converts a refund (money leaving) into an exchange (value staying) with dynamic same-value filters, turning a loss moment into a second sale.
Let users rehearse risky actions: Whatnot's interactive onboarding lets you practise "swipe to bid" with explicit "this won't charge you" context, removing the fear before the first real bid.
Group by human, not by item: Chipotle asks who each item is for and groups the basket by person, making a group order easy to double-check.
Peter Ramsey, Built for Mars
Why I picked this: With this week's headlines all about code getting cheap, Ramsey's tagline lands hard - if anyone can build the feature, the differentiation moves to these small design judgements, and this is the sharpest weekly source for them.
🎙️ Pod Shots - Bitesized Podcast Summaries
Remember, we've built an ever-growing library of our top podcast summaries (145+). 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 #145 - The Floor Is Falling
Somewhere this week, a $600,000 contract went to zero. Not renegotiated, not downgraded. Cancelled. The cost of producing things - a contract, a CRM, a therapy candidate, a working feature - has collapsed. What makes this week's nine conversations interesting is that every one celebrating the collapse was answered by another finding the floor.
The story isn't that production got cheap. It's that the scarce thing moved - and nobody agrees yet on where it moved to. Fred Turner vibe-codes away a $600k Salesforce contract while his Anthropic bill 6x's every month. SaaStr names the flip side: "stealth churn", where a customer's internal agent quietly rebuilds your useful 20% and you find out at non-renewal. Dex Horthy's team ran a "dark factory" of unread AI code for four months and paid for it with a mis-routed primary key. The floor, it turns out, holds at architecture, judgement, and the evidentiary bar - the things benchmarks can't see.
Key insights from the full round-up:
💸 A $600k contract went to zero - Turner: Curative vibe-coded its own CRM in two months and cancelled Salesforce. Contracting went from $1,500-$2,000 a deal to about $70.
🥷 Your customer won't tell you why they left - SaaStr: internal agents inspect your thin API surface, rebuild it, and do. Call it stealth churn.
📉 The model layer repriced in public - Kimi K3 at 2.8T params with open weights promised; the "us or China" framing is really closed labs versus open ones.
🏗️ The floor holds at architecture - Horthy: four months of unread AI-written code ended in a mis-routed primary key. Benchmarks can't see bad design.
🧱 Know which layer you're standing on - O'Driscoll: hyperscalers spending ~$688BN this year. Software isn't dead, it got harder.
🔥 Somebody's paying for all this - Segal's survey of ~6,000 workers: burnout up from 44.7% to 55.7% in a year.
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 🍽️.