I was recently asked how worried we should be about AI and the risk to human life. As an early stage investor, I spend much of my time thinking about what this technology makes possible. But the opportunity becomes more credible when we take the risks seriously. My concern is whether our ability to govern AI can keep pace with our ability to build it.

What “alignment” means and why it is difficult

AI alignment means ensuring that AI systems reliably act in ways consistent with human intentions and values. The challenge is that specifying a goal is much easier than specifying everything a system should and should not do to achieve it. Ask an AI agent to increase revenue, and we mean through legitimate sales, with honest claims and respect for customers. We do not mean by manipulating vulnerable people or concealing refund requests. Those constraints may seem obvious to us, but we cannot assume a model will consistently apply them, particularly in unfamiliar situations.

Modern AI systems learn behaviour through training rather than following an exhaustive set of rules written by developers. We can reward helpful answers and penalise harmful ones, but performing well in training and safety tests does not establish that a system will behave safely in every situation. As agents gain access to software, money and communications, the consequences of that gap grow. A system can be highly effective at completing a task while choosing a method its user would never authorise.

There is also a human disagreement underneath the technical problem. Whose values should a model follow when a user, an employer, a developer and a government want different things? A perfectly obedient system could still help someone cause harm. We therefore need both reliable ways to control AI and legitimate ways to decide its boundaries. Greater intelligence does not automatically provide either.

Why the warning signs deserve attention

The reasons for concern have become more concrete. In June 2025, Anthropic tested 16 major models in simulated corporate scenarios. Under deliberately constructed conditions, models sometimes chose blackmail or corporate espionage when their goals or continued operation were threatened. These were stress tests designed to expose failures, not evidence that deployed assistants routinely behave this way.

The research has continued. Anthropic’s Summer 2026 follow-up describes additional simulated failures, including covert interference with research workflows, while acknowledging progress against earlier blackmail evaluations. The researchers explicitly caution against interpreting the test frequencies as real-world failure rates. The useful lesson is that fixing a known test does not establish that a system will behave reliably in every new setting. I wrote about this on VC Cafe in “The AI didn’t go rouge, it followed a goal“.

There are also safety questions in everyday interactions. OpenAI’s May 2026 update on sensitive conversations describes efforts to detect risks that emerge gradually across a conversation. That is a different problem from an autonomous agent concealing an action, but both illustrate why evaluating an isolated answer is insufficient. Context, duration, access and the vulnerability of the person using the system matter.

This is why the move from chatbots to agents changes the stakes. A bad answer can mislead someone. A system authorised to send messages, modify software or spend money can turn a bad decision into an action before anyone reviews it. We do not need to agree on a date for superintelligence to recognise that distinction.

The money and geopolitics pushing AI forward

Dario Amodei and Sam Altman offer revealing perspectives on this tension. In The Adolescence of Technology, Amodei describes risks spanning autonomous behaviour, malicious use, authoritarian power and economic disruption. He argues for targeted intervention while acknowledging uncertainty and rejecting fatalism. His concern about China also informs his support for restrictions on access to advanced chips.

Altman’s The Gentle Singularity puts greater emphasis on abundance, adaptation and broadly distributed access to intelligence. Yet he explicitly identifies solving alignment as a prerequisite in his proposed path forward. My reading is that both see extraordinary benefits and serious risks, with different emphases on how society navigates the transition.

Their views deserve attention, and independent scrutiny. They have unusual access to the technology and lead companies with a stake in how it is regulated. A warning can be sincere while also supporting a preferred policy. A safety commitment can be meaningful while remaining vulnerable to competitive pressure. Neither observation requires assuming bad faith.

The money makes that pressure tangible. In February, Reuters reported Bridgewater’s estimate that Alphabet, Amazon, Meta and Microsoft would invest roughly $650 billion in AI-related infrastructure during 2026, compared with $410 billion in 2025. Those were estimates, not audited full-year spending, and infrastructure budgets should not be confused with spending exclusively on frontier model training.

Separately, Stargate was announced in January 2025 with an intention to invest $500 billion over four years in US infrastructure for OpenAI. Its initial funders included SoftBank, OpenAI, Oracle and MGX; its technology partners included Microsoft and Nvidia. This is a multi-year ambition, not another annual spending figure to add to the hyperscaler total.

These relationships show how deeply interconnected the incentives have become. An infrastructure provider can be a partner, supplier and beneficiary of a lab’s expansion. Investors need returns, data centres need customers, and governments want jobs and strategic capability. My concern is that, as commitments accumulate, a decision to delay a release becomes harder to make. That does not prove anyone will compromise safety. It makes independent oversight more necessary.

The same tension operates between countries. The US AI Action Plan announced in July 2025 explicitly prioritised innovation, infrastructure and international leadership, including faster data centre development and exports of American AI systems. China, meanwhile, called for a global governance framework and argued that AI should remain under human control while promoting wider access to its technology.

These stated positions illustrate why safety and influence are difficult to separate. Governments want a say in the rules and in whose systems the world adopts. If each side treats restraint as a concession to its rival, shared concern can still produce collective acceleration. National security is a legitimate consideration, but winning a technological lead cannot by itself establish that the technology is safe.

Who should hold the AI industry accountable?

One concrete proposal comes from Demis Hassabis. In his July 2026 framework for frontier AI, the Google DeepMind CEO proposed a standards body modelled on FINRA. The Financial Industry Regulatory Authority is a self-regulatory organisation that oversees US broker-dealers under the supervision of the Securities and Exchange Commission. Its relevance is the combination of industry expertise and funding with government oversight and enforceable rules.

Hassabis envisages a substantially industry-funded body with independent technical experts and open-source representatives. It would develop regularly updated safety assessments, initially review models voluntarily before release, and move towards mandatory assessments for frontier models deployed in the US once the process is proven. Tests would cover areas such as cyber and biological risks, deception and attempts to bypass safeguards. Models below the frontier thresholds would be exempt from this particular process.

I think the proposal deserves serious consideration. It gives institutional shape to the demand for independent testing, with resources to recruit technical talent. But who oversees the overseer matters. I would want independent governance, transparent decisions, meaningful enforcement and an appeals process, alongside safeguards against the largest labs using standards to exclude competitors. Industry funding should buy the capacity to scrutinise companies, without buying influence over the outcome. A US body could also help establish shared standards internationally, although securing Chinese participation would remain a separate diplomatic challenge.

So who should guard AI? Responsibility needs to be distributed, with enough independence that no single company or government gets to judge all its own decisions. A FINRA-style institution could carry part of that responsibility. I would focus on five practical priorities:

1. Test consequential capabilities independently. External evaluators need meaningful access and time to investigate dangerous behaviour before high-risk deployment. Tests should evolve as systems gain new tools and permissions, and include the complete product rather than only the underlying model.
2. Make autonomy something a system earns. Start with limited permissions, constrain spending and data access, and require approval for consequential actions. A human reviewer needs the information and time to intervene; an approval button alone is weak protection.
3. Make failures visible and assign responsibility. Serious incidents need reporting and investigation. Model developers, application builders and deploying organisations should have explicit responsibilities. A company should not be able to sell an autonomous service and then use its autonomy as an excuse when it causes harm.
4. Tie oversight to risk and preserve competition. A creative tool and an agent controlling critical infrastructure warrant different scrutiny. Rules should target demonstrable hazards and deployment conditions, with affordable routes for startups to show they meet the requirements.
5. Pursue narrow international agreements. Even rivals have an interest in preventing catastrophic misuse and accidental escalation. Shared evaluation methods, incident communication and limits on particularly dangerous uses offer concrete starting points, without requiring agreement on every aspect of AI policy.

None of this solves alignment on its own. It can reduce exposure to failures while technical research continues, and create consequences when commercial pressure overwhelms caution.

The startup opportunity in making AI trustworthy

I remain optimistic about the opportunity for startups. Cheaper access to powerful tools can help small teams build products that previously required far more capital and specialist labour. There is enormous potential in education, accessibility, scientific discovery, creative expression and helping people spend less time on administrative work. These benefits are worth pursuing with urgency.

There is also a business opportunity in making AI dependable enough to use. Customers need to know what an agent can access, whether it completed a task correctly, when it should escalate and how to recover from mistakes. Founders who solve those problems in a specific workflow can make adoption possible. They will still need differentiation as the model providers improve their own safeguards, but deep domain knowledge and evidence of reliable outcomes can matter more than a generic safety promise.

For investors, that should change the diligence conversation. Alongside asking what a product can automate, we should ask what it is authorised to do, what happens when it is wrong and who bears the cost. Reliability belongs in the product and business model from the beginning.

I do not think we should assume it is already too late. The scope of autonomy, the permissions we grant, the standards we require and the accountability we establish remain choices. AI can deliver extraordinary positive change, and startups will play an important part in making that happen. The strongest case for optimism is our willingness to do the work that makes those benefits more likely and the harms less so.

The original version of this newsletter was published on Linkedin to 8,400+ subscribers.

Firgun is the Hebrew word for taking genuine pleasure in someone else’s success. Every week since March 2020, we’ve made time to celebrate the people building, creating and moving our startup community forward. If this was forwarded to you, you can find previous editions on VC Cafe.

As we gather for Rosh Hashanah, I’m entering the new year with hope for Israel and for our community. There is something inherently optimistic about starting a company: committing years of your life to something that doesn’t exist yet, and convincing others to help you build it. Every week, this newsletter offers reminders of that belief in the future, through founders taking another leap, teams reaching a milestone and people opening doors for one another. This week brought investment in everything from enterprise AI and cybersecurity to water conservation and understanding the seabed, alongside exits built on years of patient work. My wish for the year ahead is that we have more room to create, more reasons to celebrate and more time with the people we love. May it bring health, healing, security and peace, and may we keep showing up for each other. ??? ???? ??????!

Today also marks 25 years since the September 11 attacks. We remember the lives lost, the families whose worlds changed forever, and the extraordinary courage of those who rushed to help. As this anniversary coincides with the start of Rosh Hashanah, we hold remembrance alongside hope, wishing for a year of greater compassion, healing and peace.

In the wider tech world, DeepSeek released V4.1-Flash, OpenAI introduced ChatGPT Images 2.5 and ChatGPT for Financial Services, and Meta joined the AI assistant race by launchin Muse, its personal AI agent. Apple unveiled its first foldable phone, the iPhone Duo which will be out in October for min $2,000, while Suno launched v6 with music-industry partners and Universal Music Group teamed up with ElevenLabs on a licensed AI music platform. Infrastructure spending kept climbing: Google committed €13 billion to Finland, including nuclear power, and Microsoft reportedly plans to more than triple its data-centre capacity by 2032. Mega-rounds included Mistral’s €3 billion at a €21 billion valuation, Cognition’s $2 billion at $48 billion, Harvey’s $550 million and Positron AI’s $875 million at $5 billion. In M&A, Bending Spoons agreed to buy Miro for a reported $1.36 billion (it was valued at $17.9B in 2022), Analog Devices agreed to acquire Alif Semiconductor for $1.35 billion plus potential additional payments, and Salesforce reportedly held talks to buy Listen Labs for around $2 billion. Alongside the commercial momentum, Anthropic’s report on disrupted AI misuse underscored the growing challenge of keeping increasingly capable agents under control.

QUOTE OF THE WEEK

“Hope is the belief that if we work hard enough together, we can make things better.”

Rabbi Jonathan Sacks

SPECIAL OFFER FROM OUR PARTNERS, HUBSPOT FOR STARTUPS

Get practical RevOps guidance and a personalized 1:1 GTM audit from ScaleOps, at no cost.

The session is complimentary for eligible Israeli startups that join the HubSpot for Startups program. Link to sign up: https://offers.hubspot.com/vc-cafe-gtm-session

NEW FUNDING ROUNDS

Announcements published since the September 4 edition. Cumulative funding disclosures are identified separately from individual rounds.

• Huge congratulations to Michael Kaufman and the Covenant team on emerging from stealth with more than $250 million raised across three rounds, to build lower-cost, long-range defence systems! The latest round reportedly valued the company above $1 billion!
• Mazal tov to Nir Zuk, Ehud S. and the Cylake team on securing $245 million funding round to build an AI security platform for organisations that need to retain control of their data and infrastructure!
• Kudos to Alon Geva a and the Wint Water Intelligence team on raising a $36 million Series D to detect water waste and prevent costly leaks in buildings using AI!
• Way to go Elad Schulman and team Lasso on a strategic $30M round of funding to scale Leap, is a transformer-free guardrail that runs entirely on ordinary server processors, with zero GPUs required!
• Congratulations to Shy Dekel and the Cymphony team on raising a $25 million Series A to help enterprises secure the access and actions of employees and AI agents!
• Well done to Sarah Levy and the Euno team on raising a $23 million Series A to give enterprise AI agents the organisational knowledge they need to work reliably!
• Kol hakavod to Tomer Adler and the HydroSight team on emerging from stealth with a $5 million seed round to turn marine survey data into actionable intelligence about the seabed!

EXITS

• Big congratulations to Gidi Cohen, Danny Kibel and the Bonfy.AI, a Kiteworks company team on their acquisition by Kiteworks , bringing real-time protection of sensitive information into its enterprise data security platform! Terms were undisclosed; CTech estimates the deal at tens of millions of dollars.
• Mabruk to Amir Fishman, Roy Fishman, Irad Gilboa and the Scanovate team, on the acquisition of their activities by Prytek, in a transaction valued at approximately $50 million in its first phase! A significant milestone for a bootstrapped business helping financial institutions manage digital identity and customer onboarding!

LINKS FOR YOUR BROWSER

REMAGINE VENTURES / VC CAFE / ISRAEL VC

• AI Assistants want to own the work we delegate – My look at the competition to become our personal and team AI assistants, and what it takes to earn the right to act on our behalf, including Instinct, Town and others
• Stripe’s $8 Billion OpenRouter Bet: Who Owns the Economics of AI? My analysis this week of the previously announced acquisition, and why AI usage, billing and margins are becoming inseparable.
• My impressions from the Israel Innovation day at the London Stock Exchange, the day after a low point in the relationship between the two countries…

ISRAEL

• Israel’s AI industry attracted $11.8 billion in 2025 – A report covered this week puts AI at 89% of Israeli high-tech investment last year, with nearly 3,000 AI companies active. These are 2025 investment figures, not this year’s total.
• The TLV AI 50 – A useful look at the people behind Israel’s contribution to the next wave of AI by TLV Partners
• There is $8.07 available for investments to Israeli startups – the H1 2026 investors report by IVC Data and Insights

GLOBAL

• Introducing AlphaGenome Atlas – Google DeepMind’s resource for exploring how changes in DNA could affect gene regulation.
• Anthropic’s September threat intelligence report – Concrete examples of attempted misuse and the company’s interventions, useful context for anyone building agent products.
• Investors spend 4 minutes per pitch deck on average – Papermark Startup Fundraising report. How investors actually read your pitch deck, measured across 24,541 decks, 358,672 investor views and the rounds that closed afterwards.

MEDIA OF THE WEEK

FDI In Israel: 90% from the USA – In 2025, the United States investments in Israel accounted for nearly 90% of all FDI.

Moonshot capitalism: AI rewrites the venture capital playbook (by Financial Times)

Every $100M+ funding round in Europe – 2026 YTD – by Dealroom.co (source)

Related: the European Diaspora Index. European founders have built $1.8 trillion of unicorn value outside Europe. It covers 416 unicorns with at least one founder who grew up in Europe, including Stripe, Hugging Face, Snowflake, Datadog, Miro, Grammarly, and more. by Dealroom.co

The Market map for AI assistants by Lightspeed India (source)

The physical AI market map by defy.vc (source)

According to The Economist, AI has created 1 million new jobs in the U.S in 2025

The humanoid robot value chain (source)

An incredible drone show in New York, in honour of the 2,977 victims who were murdered that day. May their memories forever be a blessing. Never forget.

Shana tova u metukah from all of us at the Remagine Ventures team ??

That’s all for this week. There’s a lot to celebrate in our “small” community! Wishing you and your loved ones a sweet, healthy and peaceful new year. Keep on creating!

Eze Vidra

The battle is heating up for general-purpose personal and work assistants: products that help an individual or team across multiple tasks and applications. An agent is the software that executes a task; an assistant is the continuing relationship through which users delegate those tasks. That distinguishes this race from narrower agents built exclusively for customer support, coding or another specific function.

The landscape includes personal and team assistants such as Instinct, Town, Lindy, Vellum, Poke and Jo; managed work agents such as Grok Bot, ChatGPT agent from OpenAI, Claude Cowork and Yutori Scouts; and open-source systems such as OpenClaw and NanoClaw. These categories overlap, but they help distinguish what users are buying and how they delegate work.

Why does this matter? Because the assistant could become the place where we decide what happens next. If it researches a purchase, chooses a service and completes a transaction, it can influence where money flows. If it coordinates work across business applications, it can influence which software people open, which subscriptions they value and which interfaces become less important.

That possibility helps explain the capital flowing into the category. Instinct reportedly raised a $250 million Series B at a $2.5 billion valuation, co-led by Index Ventures and Benchmark. But a valuable category does not guarantee that every assistant becomes a valuable company.

Harry Stebbings’ 20VC conversation with Town co-founder and CEO Jean-Denis Greze, and the accompanying takeaways, frame the central investment question: what can an independent assistant own as the underlying models and competing platforms improve?

Defining the competitive field

Comparing Gemini with Grok Bot is not quite apples to apples. One is a broad assistant product family; the other is a specific product designed to execute delegated work. A more useful comparison looks at the job each product takes over, how users interact with it and who operates the system.

The following landscape reflects documented positioning and capabilities, rather than a ranking based on hands-on performance testing.

The boundaries are porous. Town can execute tasks, Grok Bot can retain context, and an open-source system can support an ongoing personal relationship. The grouping describes their primary product approach, rather than exclusive capabilities.

Gemini supports recurring scheduled actions, while Meta AI already operates inside WhatsApp and other familiar interfaces. Both belong in the competitive picture. However, that presence should not automatically be equated with the execution capabilities of a dedicated work-agent product. Features also vary by plan, region and configuration.

Several bets on the same opportunity

Instinct’s bet is that personal delegation should feel as simple as contacting another person. Town places greater emphasis on professional routines and shared work: its site describes meeting briefings, inbox handling, recurring account updates and team integrations. Lindy similarly positions itself around connected tools and company context. Their potential advantage is learning how someone works and becoming useful without repeatedly asking them to explain it.

Vellum and Jo add other approaches to persistent personal assistance. Vellum now describes itself as personal intelligence, expanding beyond the enterprise AI development positioning described in its 2025 Series A announcement. Jo’s YC profile describes an agent spanning a user’s Mac and a private cloud machine. These products compete for the same underlying permission: becoming the assistant a person trusts with context and continuing work.

Poke illustrates both the opportunity and the consolidation already underway. General Catalyst led its $15 million seed round, backing an assistant delivered through familiar messaging channels. Its website now confirms it is joining Cognition. It remains relevant to the product landscape, but should no longer be presented as an independent startup.

Grok Bot takes a more explicit AI-worker approach. Its bots have persistent computers, use applications and can collaborate. ChatGPT agent and Claude Cowork bring delegated execution into products with existing users and subscriptions. Yutori Scouts approaches the opportunity through ongoing web tasks. These products put pressure on independent assistants to show why their particular experience is worth adopting.

Where OpenClaw and NanoClaw fit

OpenClaw and NanoClaw offer a different ownership model. Users can operate and customise the assistant system themselves, rather than relying entirely on a managed product. Neither is a foundation model. They organise how models use tools, maintain context and carry out work.

OpenClaw emphasises extensibility across tools and communication channels. NanoClaw emphasises a smaller, more inspectable system with container isolation. Its V2 announcement adds human approvals and persistent communication between agents.

The distinction matters because greater control comes with operational responsibility. Someone has to manage deployment, integrations, credentials and permissions. Self-hosting also does not automatically keep all processing local: external models and connected services can still receive data.

Container isolation limits what an agent can access; it does not guarantee that an authorised action is sensible. Sending the wrong email can still be damaging even if the software never escapes its sandbox. A dependable assistant needs both technical boundaries and good decisions about when to act, ask or stop.

These projects make the assistant layer easier to experiment with and reproduce. That expands the market while putting pressure on startups whose differentiation is largely a collection of integrations.

Why pay for another assistant?

This is the question every independent founder needs to answer. The major platforms already combine strong models with tools and increasingly proactive capabilities. Their users have existing subscriptions, familiar interfaces and accumulated context. An independent product must deliver a meaningful improvement over that baseline.

Consider a founder preparing for a customer meeting. Producing a briefing is useful. Consistently finding outstanding commitments, preparing the briefing, drafting the follow-up and updating the relevant systems after approval is a more valuable service. The differentiation lies in completing that process reliably, with less instruction each time.

For Instinct and Poke, the opportunity is making everyday administration easy enough to become habitual. For Town and Lindy, it is embedding assistance into professional and team routines. For Grok Bot and other managed agents, it is making delegated execution dependable. For OpenClaw and NanoClaw, control and customisation provide another reason to choose them.

The test across all of them is the same: how much work can I hand over without creating another job supervising it?

Distribution earns attention. Reliability earns delegation.

The giants have substantial advantages. Grok has a route to users through X. Google has existing productivity relationships. Meta already offers AI inside WhatsApp. Independent startups must persuade people to establish another relationship and connect another product to their working lives.

But access to users is different from permission to act for them. Someone may happily ask an assistant for restaurant suggestions while hesitating to let it contact a customer, make a purchase or change a booking. Each step towards autonomy requires the product to earn more trust.

The 20VC discussion raises doubts about whether assistants have achieved deep product-market fit beyond power users. I would frame this more narrowly: enthusiasm for AI is established, but habitual delegation remains a distinct product challenge. A successful demonstration does not tell us whether ordinary users return, pay and gradually need to intervene less.

A useful assistant should also reduce the burden of deciding what to delegate. If users must become expert prompt writers, integration managers and quality controllers, much of the promised productivity gain disappears.

Can assistants create network effects?

Greze’s emphasis on collaboration suggests a possible source of defensibility. An assistant could become more useful when it can coordinate with colleagues’ assistants, carrying context across departments and reducing the need to chase updates. Both Town’s team offering and NanoClaw’s agent-to-agent communication illustrate product development in this direction, although that alone does not establish network effects.

Imagine a salesperson’s assistant checking a customer’s commitments with the account team’s assistant before requesting a pricing exception from finance. More participating colleagues could make the system more valuable, provided it respects each person’s permissions.

But workflow lock-in is not automatically a network effect. Shared routines and organisational memory create switching costs. A genuine network effect requires each additional participant to improve the product for existing participants. Interoperability could also allow that value to accrue across competing assistant providers.

That is why I would not draw a permanent boundary between consumer and enterprise assistants. Instinct and Town may start with different users and buying decisions, but people’s personal and professional lives overlap. Their initial positioning explains their strategies; it does not guarantee they stay out of one another’s markets.

The investment case depends on completed work

The economics deserve as much attention as the product experience. A more active assistant can generate more inference, execution and support costs. Increased usage is encouraging only if the business can afford to serve it.

I would want to see cost per successfully completed workflow, including retries and human intervention, alongside paid retention and expansion. Task counts alone can conceal an expensive service. Equally, an assistant that appears cheap may deliver little value if the user repeatedly has to repair its work.

There may be several winners: a personal assistant, an assistant embedded in a productivity suite, a team-work platform and infrastructure supporting privately operated agents. The common requirement will be turning impressive capabilities into dependable delegation.

The company that earns the right to act repeatedly on our behalf could become one of the most valuable relationships in software.

That is the problem at the heart of Stripe’s agreement to acquire OpenRouter: how do you turn growing demand for AI into a business whose economics improve as it scales? The companies did not disclose financial terms, but the Financial Times reported an approximately $8 billion transaction, while Reuters put it slightly above $8 billion.

My reading is that Stripe wants to connect the cost of delivering AI with the revenue earned from it. For founders, the acquisition is a signal that model selection, product pricing and gross margins are becoming inseparable decisions.

OpenRouter helps developers access different AI models through a common interface. Instead of maintaining separate integrations with each provider, companies can use its gateway to manage model access, costs and routing. At the announcement, OpenRouter said it processed more than 10 trillion tokens a day across over 400 models, serving a community of more than 10 million developers and companies. Stripe’s announcement puts the provider count above 80.

The appeal is straightforward. Different tasks need different combinations of intelligence, speed, reliability and price. A simple classification task may not require the model used for a difficult reasoning problem. A production application also needs alternatives when a provider becomes unavailable. The gateway sits where those choices become actual consumption.

Stripe had already started moving into this territory. Its Token Billing product connects model consumption with customer billing, including tracking underlying token prices. In November 2025, Stripe described an LLM proxy that records usage, applies a company’s markup and supports invoicing. It then completed its acquisition of Metronome in January 2026, adding infrastructure for complex usage-based billing.

Stripe’s leaked shareholder letter puts these acquisitions into a broader strategic framework. The company describes agents becoming economic actors in their own right: discovering services, registering for them, consuming resources and making payments. It groups its infrastructure around discovery and onboarding, usage management, payments and fund storage. Metronome supports usage management, Bridge supports stablecoin payments, and Privy provides wallet infrastructure. OpenRouter adds access to the models that power the work.

Read together, these moves suggest an ambition to support the commercial lifecycle of an AI agent. Stripe could help an agent find a service, access a model, account for consumption and settle the resulting payment. The letter also says these capabilities will be integrated deeply into Stripe’s existing products. That gives the strategy a distribution advantage: businesses could adopt infrastructure for the AI economy through financial tools they already use.

OpenRouter therefore expands Stripe’s ambition upstream. Stripe can potentially help a business choose where to spend on inference, measure what it consumes and charge its customers. That is a strategic opportunity created by the combination, rather than evidence that every part already works as one integrated product.

The distinction matters because measuring expenditure is only half the problem. Understanding whether that expenditure creates value is harder.

AI costs already appear in company accounts. What is changing is how directly an individual product decision can affect the economics of serving a customer. Give an agent permission to reason for longer, search more sources or retry a failed task, and the cost of completing that workflow can change substantially. Engineering choices become margin choices.

This makes AI economics a shared responsibility for engineering, product and finance. The CFO should help establish budgets and acceptable margins; product teams need to understand what customers value; engineers need to determine which models and workflows deliver it reliably. At an early-stage startup, those responsibilities may sit with the founders rather than three separate departments.

The useful question is: what does it cost us to deliver a successful customer outcome? Knowing the monthly token bill is a starting point. Knowing which customers, features and workflows create profitable demand makes that information actionable.

Take for example an illustrative customer-support product. One model costs two cents per attempt but frequently requires retries or human intervention. Another costs six cents and resolves more cases correctly. Choosing the cheaper model could increase the total cost per resolution. Token prices matter, but so do failure rates, latency, tool charges and the human work left behind.

That is also where the “tokens are the new dollar” analogy reaches its limits. Tokens are units of model consumption, with prices and usefulness that vary across models and tasks. They do not measure customer value. A business can consume more tokens because it is doing more useful work, or because its agents are inefficient.

The usage data illustrates why this distinction is becoming urgent. An a16z chart based on OpenRouter data shows agent token usage overtaking human usage in February 2026 and reaching a seven-day average of approximately 7.3 trillion by early August. A separate OpenRouter spending chart for August 1–19 puts coding at 31.6% of spending and agent tasks at 28.7%, with workflow execution alone accounting for 19%.

These are snapshots of activity on OpenRouter, rather than the entire AI market. But they show a platform where substantial consumption is tied to software execution and automated workflows. One user instruction can initiate many model calls, making the economics beneath a simple interface increasingly important.

For Stripe, occupying that point in the workflow could be valuable. Payments infrastructure participates when money changes hands. A model gateway participates when the application performs work, giving it an opportunity to improve how resources are selected and consumed.

Still, sitting in the token flow does not confer unlimited pricing power. Model providers set their own prices, large customers can negotiate directly, and developers can use competing gateways. Stripe’s opportunity depends on making the combined service sufficiently useful that customers choose to keep it in their infrastructure.

The shareholder letter helps explain why Stripe sees AI as an extension of its existing business. The company argues that the qualities developers value, including programmability and frictionless setup, are equally valuable to coding tools and agents. Its approach to serving developers could therefore become an advantage as software increasingly selects and operates other software. The strategic bet is that infrastructure built to help people start and run internet businesses can also serve businesses where agents perform more of the work.

For founders, the acquisition raises three practical questions:

1. Does greater usage make your business more profitable? Revisit the economics of your heaviest users. Model selection, usage allowances and pricing need to work together so that your best customers remain good business.
2. What do you own as the infrastructure consolidates? As routing and billing become easier to buy, differentiation needs to come from the workflow, customer relationships and feedback that improve your product. Infrastructure startups face a sharper version of that challenge: why will customers choose a standalone tool over Stripe’s expanding platform? Keep alternatives available for critical dependencies.
3. Can an agent become your customer? Stripe’s shareholder letter points to a new distribution opportunity: software discovering, evaluating and purchasing other software. For API and software businesses, clear documentation, transparent pricing and onboarding that agents can complete with appropriate permissions could become part of the sales funnel.

As an early-stage investor, I see OpenRouter as evidence that substantial value can accrue to companies helping customers use models effectively. But Stripe’s expanding role also raises the bar for what remains defensible above that infrastructure.

Stripe wants to help more businesses participate in the AI economy while supplying more of the systems they depend on. For founders, the opportunity is to use those systems to build faster while retaining ownership of the customer relationship and the value delivered. Stripe can help meter and monetise the work. The startup still has to make that work worth paying for.

The original version of this newsletter was published on LinkedIn to 8,400 subscribers.

Firgun (pronounced ‘feer-goon’) is an Israeli term describing genuine, unselfish joy in someone else’s success… It’s basically the opposite of “schadenfreude”. Every week since March 2020, we celebrate the wins of our startup community. Read more about why Firgun matters. If this was forwarded to you and you’d like to subscribe, previous editions are on VC Cafe.

There was a lot to celebrate in our “small” community this week. Israeli startups announced approximately $1.4 billion in funding in the first three days of September, more than twice the $577.5 million raised during the entire month of August. Wonderful, Upwind Security Lyte and HiBob accounted for more than $1 billion between them, but the breadth was equally encouraging: capital also went into AI security, defence technology, self-healing infrastructure, consumer cybersecurity, smart toys and executive productivity.

In the wider tech world, OpenAI launched GPT-6 Astra, its first model to reach a “Critical” cybersecurity threshold, while Anthropic released Claude Fable 5.1 and Mythos 5.1, Google introduced Gemini 3.8 Flash Cyber, Meta launched Muse Voice Transcribe, World Labs unveiled its Atlas 3D world model and Microsoft introduced MAI-Transcribe-2. In M&A, Palo Alto Networks reportedly paid $500 million for Console and ServiceNow reportedly agreed to acquire Israeli startup Sweep for hundreds of millions. Mega-rounds continued as Crusoe reportedly raised more than $3 billion at a $30 billion valuation, Thinking Machines Lab discussed raising $5 billion to $6 billion at more than $40 billion, Cognition reportedly sought $1 billion at $47 billion, Fluidstack raised $1.5 billion at $18 billion and AfterQuery reached a reported $3.2 billion valuation. Finally, John Ternus succeeded Tim Cook as CEO of Apple , Uber and Wayve launched London’s first commercial robotaxi service, Google introduced the Gemini-powered Google Pics design suite, and the FTC and 22 US states sued Amazon over alleged hidden advertising surcharges.

QUOTE OF THE WEEK

“Glory is being happy. The glory is not winning here or winning there.”

Rafael Nadal

SPECIAL OFFER FROM OUR PARTNERS, HUBSPOT FOR STARTUPS

Get practical RevOps guidance and a personalized 1:1 GTM audit from ScaleOps, at no cost.

The session is complimentary for eligible Israeli startups that join the HubSpot for Startups program. Link to sign up: https://offers.hubspot.com/vc-cafe-gtm-session

NEW FUNDING ROUNDS

• Huge congratulations to Bar Winkler, Roey Lalazar and the Wonderful team on raising a $550 million Series C at a $5 billion valuation to create the AI operating system for agents in the enterprise!
• Congrats to Amiram Shachar and the Upwind Security team on raising $300 million at a $3.8 billion valuation to help enterprises identify and prioritise the cloud threats that pose the greatest risk!
• Mazal tov to Ronni Zehavi and the HiBob team on raising $166 million at a $3.2 billion valuation to help organisations manage their people, workforce data and, increasingly, AI agents through its Bob platform!
• Kudos to Alexander Shpunt and the Lyte team on raising $165 million at a $1.6 billion valuation to develop advanced perception technology that helps autonomous machines understand and interact with the physical world!
• Congratulations to Yair Saban the Air team on emerging from stealth with $50 million to build a preventative security layer that protects enterprises from malicious tools and extensions connected to AI agents!
• Well done to Udi Oster and the AITAN team on raising $41 million to take its combat-tested autonomous drone systems and edge-AI technology to Israel’s allies around the world!
• Great milestone Amos P. and the Guardio team on raising $40 million at a $1.1 billion valuation to protect consumers from increasingly sophisticated online scams, identity theft and AI-powered cyber threats!
• Mabruk to Itai Gafni and the Huskeys team on raising $27 million to bring an agentic AI layer to web application firewalls and help enterprises defend against attacks initiated by autonomous agents!
• Awesome news Shahar Abuhazira and team Easy Aerial Inc. on securing a $20M series B to scale their tethered drones for aerial surveillance and defense!
• Kol hakavod to Udi D. and the Particula team on raising $15 million to expand its portfolio of connected physical games, including GoCube, GoChess and licensed products based on major entertainment franchises!
• Congratulations to Ishay Yaari and the DataAgent team on emerging from stealth with a $10 million pre-seed round to build AI agents that detect and autonomously repair failures inside cloud infrastructure!
• Good stuff Haran Cohen Hillel and team Newlight on securing a $9M seed round to power large commercial vessels with hydrogen!
• Huge congrats to Nir Sabato and the Catch team on raising a $5 million seed round to build an AI executive assistant that autonomously manages scheduling, travel, correspondence and other administrative work!
• Big thumbs up Michael Assraf and team Flamingo on coming out of stealth with $4.5M in seed funding to build an open, AI-native infrastructure layer for IT and security serving MSPs!

EXITS

• Big congrats Ido Gaver and team Sweep on the acquisition by ServiceNow in a deal estimated at hundreds of millions to build AI agents that can work across enterprise platforms including ServiceNow, Salesforce and HubSpot!
• Well done to Ruti Ben-Shlomi and team LightSolver on its acquisition by Nasdaq-listed CollPlant for an undisclosed sum, providing a public-market route for its laser-based computing technology!
• Congrats to Aviv Shapira the XTEND team on beginning trading on the NYSE following its merger with @JFB, at an opening valuation of approximately $1.5 billion!

NEW FUNDS

• Congratulations Omri Casspi, Ori Striechman and team Swish Ventures on securing a new $250M fund to invest in 12 or so Israeli startups in cyber, AI, defense and infrastructure!

LINKS FOR YOUR BROWSER

REMAGINE VENTURES/ VC CAFE/ ISRAEL VC

• Wonderful’s $5B Bet to Own the Enterprise AI Layer – Wonderful’s real opportunity is not simply connecting agents, but owning the operational layer that turns unpredictable models into governed, measurable enterprise work.
• Why Israeli Unicorns Are Built in America – Seventy percent of Israeli-founded unicorns were built in the US. Israel can remain the engine room, but America increasingly needs to be part of the original design.
• The Right Model for the Job – Model routing could reduce inference costs and improve performance, but it also creates a powerful new control point in the AI stack.

ISRAEL

• Israeli startups raised $1.16 billion in the first two days of September – Nine rounds made the opening of September one of the busiest funding periods in recent Israeli tech history.
• Anduril appoints former Israeli Air Force commander Amiram Norkin to lead its Israel activity – A significant vote of confidence in Israel’s defence-technology ecosystem.
• Venture capitalist Oren Zeev’s sharp warning: why Israeli startups must incorporate in the US – “I’ve never met a founder who regretted a US incorporation, and I often meet founders who regret an Israeli incorporation”, claims Zeev against the common practice of startup incorporation in Israel

GLOBAL

• The State of Enterprise Software 2026 – SVB examines AI’s impact on software funding, pricing, M&A and the shift from seat-based subscriptions to usage and outcomes.
• The State of AI in the Enterprise – Deloitte finds that many organisations are using AI for surface-level productivity, while a smaller group is redesigning products and core processes around it.
• AI capital expenditure could exceed the cost of railways and the internet combined – PwC estimates cumulative AI investment could approach $32 trillion by 2050.
• Bonus: The current state of humanoid robotics – A grounded assessment of the generalisation, reliability and long-task challenges that robotics still needs to solve.

MEDIA OF THE WEEK

An interesting, if perhaps unsurprising finding, by Dealroom.co -70% of unicorns created by founders from Israel were built in the US – I went deeper in my VC Cafe post on ‘why Israeli unicorns are built in America‘.

It was a huge week for Israeli startups, so I naturally made a small infographic in the spirit of #FIRGUN – funnily enough, 30 mins later it was already obsolete as new rounds were announced! (source)

All the Israeli startups that raised funding in August and the funds who backed them, as published on Israel VC this week! (source)

According to the H1/2026 Tech Review by IVC and LeumiTech, multinational companies continue to expand their R&D presence in Israel. During the first half of 2026, 15 new multinational centers opened, while 8 closed, leaving a positive net balance of 7. This continues a pattern visible throughout the past decade, with openings consistently outnumbering closures despite considerable fluctuations in activity. Israel is currently home to 575 active multinational R&D centres. by IVC Data and Insights

AI security is very hot right now. Here are the companies that are on the M&A radar, according to CB Insights (source)

European AI category leaders according to Headline (source)

2026 has already produced more European exit value than 2024 and 2025 combined: $65B so far, compared with $52B across both full years. Source: Dealroom.co

Finally, why is Chinese open-source AI a threat to American AI companies… by Bloomberg News

That’s all for this week. There’s a lot to celebrate in our “small” community! Keep on creating!

Eze Vidr

That gap helps explain why Israeli-Dutch startup Wonderful raised $550 million yesterday at a $5 billion valuation, more than double its valuation six months ago. The company is often described as an AI-agent platform, but investors are underwriting a much bigger ambition: becoming the operational layer that turns unpredictable models into governed, measurable enterprise work.

Wonderful, founded in 2025 by Bar Winkler and Roey Lalazar, has now raised more than $800 million to date. As reported by The Wall Street Journal, Insight Partners led the Series C with a $150 million investment, while Salesforce joined existing backers Index Ventures, IVP, Vine Ventures, 9Yards and Bessemer Venture Partners. In little more than a year, Wonderful has expanded into more than 35 markets, grown to approximately 650 employees and reportedly increased its annual revenue run rate from around $1 million to $70 million.

Those figures are extraordinary, but they do not yet prove that Wonderful has built an enduring software company. Its reported gross margin is approximately 52%, reflecting a business that still relies heavily on local teams and forward-deployed engineers. The $5 billion bet is that Wonderful can turn this labour-intensive deployment advantage into a scalable, independent software layer before foundation-model providers and incumbent SaaS companies absorb the opportunity.

To succeed, Wonderful must prove three things: that forward deployment can become repeatable software, that its orchestration layer retains value as agent protocols become standardised and that it can own the enterprise control plane before model providers and systems of record claim it.

1. Can deployment become repeatable software?

Wonderful’s primary value proposition today is not simply that it builds capable agents. It gets those agents into production.

Access to intelligence is no longer particularly scarce. Enterprises can choose between models from OpenAI, Anthropic, Google, Meta and a growing selection of open-source providers. What remains scarce is the ability to connect those models to proprietary data, permissions, legacy systems and real business processes, then make the resulting system reliable enough to act on behalf of the organisation.

Wonderful addresses this through forward-deployed engineers who work directly with customers, sometimes on-site. They map workflows, connect internal systems, establish tests and guardrails, monitor agents in production and continuously improve their performance. Implementation is not something that happens after the software has been sold. It is central to the product.

Wonderful’s strategic collaboration with McKinsey makes this explicit. The partnership combines Wonderful’s platform and forward-deployed engineers with McKinsey and QuantumBlack’s transformation and change-management capabilities. McKinsey says that while 79% of organisations are experimenting with generative AI, fewer than 10% have scaled AI agents. The bottleneck is not ambition or access to models. It is execution.

Wonderful’s geographic strategy compounds that deployment advantage. While Sierra and Decagon concentrated heavily on the US customer-service market, Wonderful built its beachhead across Europe, Latin America, Asia and the Middle East. The company told Geektime that it deliberately focuses on languages other than English and Mandarin, where the market is already crowded with large players.

This was more than geographic arbitrage. Language, regulation, local culture and legacy infrastructure all affect whether an agent works in production. In a Google Cloud case study, Wonderful explains that producing a voice agent for Italy can require multiple variations to account for dialects, gender and regional differences. A generic English-language agent cannot simply be translated and expected to deliver the same performance.

Customer service also gave Wonderful a strong initial wedge. It contains a large volume of repetitive work, has clear human benchmarks and offers an ROI that can be measured quickly. Once Wonderful connects to an organisation’s systems and learns how its processes operate, it can expand into adjacent workflows.

At Colombia’s Banco Caja Social, Wonderful says it took a collections agent from zero to production in 19 days. Within six weeks, the agent completed more than 43,000 calls, increased contact-to-promise conversion from 45% to 65% and reduced average handling time by 33%. The bank then used the same foundation to launch a second agent for inbound customer service.

At Telefónica, Wonderful reports that a billing agent resolves 77% of the relevant issues it handles. Other deployments include Petrol Ofisi, where the company says it reduced IT call-handling time by 75% in three weeks, and ELTA Systems, where an internal service agent reportedly eliminated peak waiting times and quadrupled call capacity within five weeks. These are company-published results rather than independently audited figures, but they show the commercial playbook Wonderful is trying to repeat.

Its Israeli customers reportedly include Bezeq, Bank Hapoalim and Paz, while Deutsche Telekom is among its international clients. The pattern is consistent: identify a constrained workflow, reach production quickly, demonstrate a measurable result and use the initial integration as a bridge into more of the enterprise.

The services-heavy model should not automatically be viewed as a weakness. In “Trading Margin for Moat,” Andreessen Horowitz partner Joe Schmidt argues that “sometimes human-intensive services are necessary to create transformative software.” Salesforce, ServiceNow and Workday all required significant implementation work before developing scalable platforms and partner ecosystems.

As I argued recently in “The $9 Billion Bet on Forward-Deployed Engineers,” fieldwork only becomes a moat when it becomes product. If every engagement starts from zero, the startup is building a consultancy. If each deployment produces reusable integrations, evaluation tools and knowledge that make the next one faster, the forward-deployed team becomes a product engine.

This is the first test for Wonderful. The important metric is not how many engineers it can place with customers, but whether the time, labour and cost required for each subsequent deployment decline.

2. Does orchestration remain valuable as connectivity becomes standardised?

As enterprises deploy more agents, they risk recreating the fragmentation of the SaaS era. Sales may adopt one agent, customer support another and finance a third, each connected to different models, data sources, applications and permissions.

Consider a customer requesting a refund. One agent might interpret the request, another retrieve the customer’s history, another review the relevant policy, another authorise the transaction and another update the CRM. The system must determine which agent should act, what information it can access, which model is appropriate and when a human needs to intervene.

Wonderful wants to provide that control layer. Its AI operating system is designed to coordinate agents, workflows, enterprise context, integrations and governed execution. Customers can choose different models for different tasks and deploy the platform in the cloud or on-premise.

Wonderful also says customers can export their agents, skills, tools and governance configurations, run external agents inside Wonderful or use Wonderful agents elsewhere. This “open by default” approach matters because enterprises will be reluctant to hand a critical operating layer to a platform that locks them into one model provider.

However, orchestration is important without being inherently defensible. Open standards such as Model Context Protocol and the Agent2Agent Protocol are standardising how agents connect to data, tools and other agents. Basic agent routing and connectivity could therefore become commodities.

The infrastructure opportunity is not merely connecting agents. It is owning the operational layer that turns unpredictable models into governed, measurable enterprise work.

That layer includes enterprise context, identity, permissions, evaluations, monitoring, exception handling and accountability. A company needs to know not only that an agent completed a task, but why it took a particular action, which information it accessed, whether it followed company policy and who remains responsible when something goes wrong.

Wonderful’s testing and improvement tools could become particularly valuable here. According to Geektime, its platform can create tests, automate aspects of prompt optimisation, monitor performance and use AI to analyse interactions. A platform observing millions of real conversations and workflows could build proprietary evaluation datasets that generic benchmarks cannot reproduce.

This is where the analogy with Cursor becomes relevant. Cursor initially created value on top of models from OpenAI and Anthropic. Its advantage came from controlling the developer workflow and designing the interface, context and tools around those models.

Once Cursor owned that workflow and accumulated data about how developers use AI to search, write, edit, run and debug code, it moved deeper into the stack. It introduced Composer, its own specialised agentic coding model, trained through reinforcement learning on software-engineering tasks. Cursor still supports third-party frontier models, but can now route work between external and proprietary models.

Wonderful could follow a similar path. Its advantage today is not a proprietary foundation model, but its proximity to production workflows and the data created while deploying and improving agents inside enterprises. Over time, it could use those interactions to build specialised evaluation systems, fine-tune models for particular industries and languages or train proprietary models where doing so improves performance or economics.

An application company does not need to own the entire stack from the beginning. It can first own the workflow and customer relationship, then move down the stack where its proprietary data gives it a right to win.

The second test for Wonderful is therefore whether it can build durable value above the connectivity layer. Its moat cannot be the ability to pass tasks between agents. It must come from understanding how the enterprise operates and proving that AI completed the work correctly.

3. Can Wonderful own the control plane independently?

Wonderful’s long-term competitors are not limited to Sierra and Decagon. Model providers including OpenAI, Anthropic and Google are moving into enterprise applications and deployment. Systems of record including Salesforce, Microsoft and ServiceNow already control valuable enterprise data and workflows, while consultancies such as McKinsey, Accenture and Deloitte own transformation budgets and senior customer relationships.

This raises a fundamental question: why should a separate orchestration company own the enterprise AI control plane?

The foundation-model companies can combine orchestration with proprietary intelligence. Salesforce can build agents around customer data already stored in its platform. ServiceNow can coordinate agents around established IT and enterprise workflows, while Microsoft can distribute agents through Office, Teams, Dynamics and Azure.

Wonderful’s strongest answer is neutrality. It can sit above different models and enterprise applications, allowing customers to select the best technology for each workload without replacing their existing stack. As models improve and prices fall, a neutral platform should theoretically benefit rather than be disrupted.

Salesforce’s investment in Wonderful is therefore both validation and a source of tension. Salesforce wants Agentforce to coordinate work around the customer data it already controls, while Wonderful wants to sit across multiple models and enterprise applications. The investment could represent a partnership, a hedge or an option on a company that might otherwise become a strategic threat.

Wonderful’s international footprint may give it time to establish this position. Local engineers, language expertise, regulatory familiarity and relationships with large enterprises are harder to reproduce than an agent demo. By avoiding a head-on fight for the same early US customers, Wonderful has built a credible alternative while the category is still forming.

Its extraordinary funding also contributes to that advantage. A $550 million round signals to customers, employees and partners that Wonderful is likely to remain in the market. Financial durability matters when a CIO is choosing infrastructure that may eventually coordinate hundreds of business-critical agents.

Capital can also crowd out competitors. A smaller startup may have strong technology, but customers will question whether it can match Wonderful’s geographic coverage, local deployment capacity and balance sheet. Employees may prefer the perceived category leader, while investors may hesitate to fund another company directly in its path.

This is venture capital as kingmaking. Insight Partners has now led three consecutive Wonderful rounds and describes its latest investment as “tripling down” on the company. The objective is not only to finance growth, but to establish Wonderful as the default independent platform before the market fully settles.

Capital, however, is not the same as defensibility. A large balance sheet can subsidise bespoke deployments, aggressive hiring and geographic expansion for longer than a smaller competitor could afford. It can also conceal whether the underlying business is becoming more efficient.

Wonderful’s reported 52% gross margin and approximately $108,000 in annualised revenue per employee do not yet resemble mature software economics. At a $5 billion valuation, the company is valued at more than 70 times its reported revenue run rate. Investors are underwriting a future in which its deployment work becomes more repeatable, customers expand into multiple workflows and gross margins rise as implementation knowledge is converted into software.

There are historical precedents. Andreessen Horowitz notes that ServiceNow had a gross margin of 63.2% at IPO and Workday’s was 54.1%, well below the margins they later achieved. Salesforce also invested heavily in implementation before developing the partner ecosystem that helped it scale.

The third test is whether Wonderful can make the same transition while remaining an independent platform. It must become sufficiently embedded to be indispensable, yet sufficiently open to avoid being absorbed into the ecosystems of the companies beneath and around it.

The real $5 billion bet

Wonderful has already proved that there is enormous demand for getting AI agents out of pilot mode and into production. It has also demonstrated that an international, deployment-led strategy can create a meaningful alternative to competing head-on for the same US customers.

What it has not yet proved is whether that deployment machine can become repeatable software, whether its orchestration layer will retain value as agent protocols become standardised and whether an independent company can own the enterprise control plane before model providers and systems of record absorb it.

That is the real $5 billion bet. Wonderful is not merely trying to build better agents or connect them together. It is trying to own the operational layer that makes enterprise AI reliable, governed and economically useful.

Deployment has given Wonderful a credible head start. The question is whether it can turn that head start into infrastructure.