The first four conversations were recorded live for ScaleUp Nation, hosted by Erez Shachar and Ram Ben Ishay. The podcast aims to become an Israeli counterpart to Acquired (one of my personal faves), exploring the stories and decisions behind the country’s leading technology companies and how they grow into global businesses.

Listening to the presentations and to the founders behind HiBob and Wiz (now part of Google), I found myself thinking about how much of our understanding of scale is tied to the way software companies used to grow. We count employees, celebrate hiring and associate larger teams with greater ambition. But what happens when a small team can build a business that previously required hundreds of people?

Survival and scale require different skills

One of the most useful distinctions in the presentation is between surviving and growing. Founder relationships, trust and a team’s technical foundations can help it navigate the early years. The slides associate growth among surviving companies with experience in large organisations and access to international networks.

These are associations, rather than proof that a particular background causes success. But the implication deserves attention. Building a global company requires familiarity with how customers buy, how larger organisations operate and how products behave when usage expands.

The founder who can personally rescue every customer relationship eventually needs a team that can deliver without that intervention. An improvised sales process needs to become repeatable. Engineering decisions must support reliability, security and cost at a much larger scale.

This gives Israel’s multinational R&D centres an additional role. They can develop future founders and managers who understand global operations. But that benefit grows when people gain exposure to customers, product ownership and commercial decisions alongside engineering.

The slides describe a useful bridge: bringing experienced global operators into growing companies. The goal should be to match experience to the next constraint. A company still searching for repeatable demand needs a different executive from one expanding an established sales organisation.

Growing globally and retaining value locally

International expansion is essential for companies serving markets much larger than Israel. The concern begins when the centre of expertise and decision-making moves with it.

The Israel Innovation Authority’s 2026 high-tech report found that the share of employees based in Israel at private Israeli high-tech companies fell from 69% in January 2019 to 62% in March 2026. It also highlighted the expansion of management and R&D activity abroad.

In an October interview with Globes, Innovation Authority chairman Alon Stopel cited approximately 440,000 people employed overseas by Israeli companies, including public companies, compared with approximately 425,000 high-tech employees in Israel in 2025. These are different populations, so the figures should not be treated as a precise split of one workforce. They nevertheless illustrate the international footprint of Israeli technology.

There are legitimate business reasons for hiring near customers and recruiting talent abroad. For Israel, the challenge is to remain an attractive location for the functions that shape the company’s future: leadership, product strategy, advanced research and the development of the next generation of managers.

Trust matters here. A July 2023 Startup Nation Central survey, also cited in the slides, found that 68% of surveyed startups had taken legal or financial steps in response to judicial uncertainty. Those actions included moving cash, relocating headquarters or employees, and layoffs. It was a broad measure taken at a specific moment, rather than evidence that 68% had left Israel.

The policy lesson remains relevant. Stable institutions, predictable rules and confidence in the future influence where founders build their companies and families. Resilience helps people withstand uncertainty; public policy should help reduce it.

What does scale mean in the age of AI?

One question from the day stayed with me: if AI allows a company to grow revenue dramatically without growing its team, what should we count as scale?

The 2007 Eurostat–OECD Manual on Business Demography Statistics, cited in the presentation, defines high-growth enterprises using “average annualised growth greater than 20% per annum, over a three year period”. Growth can be measured by employment or turnover. The manual proposes an initial threshold of ten employees for both measures.

The distinctions matter: this is a definition of high-growth enterprises, often used in discussions of scale-ups; it measures average annualised growth over three years, and does not require both revenue and headcount to increase. More recent Eurostat methodology uses a threshold above 10% for its employment-based high-growth measure.

The revenue-based definition can already accommodate a company growing with a small team. What AI challenges is our habit of treating hiring as evidence of progress. A company that doubles revenue while keeping its team the same size may be scaling exceptionally well.

For me, that shifts the investor conversation towards the quality of growth. Revenue per employee is useful, but so are retention, gross margins after inference and delivery costs, and how much human intervention each additional customer requires. A small team can still be running an expensive, labour-intensive business behind the scenes.

For Israel, this creates both opportunity and tension. More small teams could build significant global businesses. Yet the economic success of those businesses may create fewer direct jobs than previous generations of software companies. We need to distinguish company productivity from the broader challenge of creating opportunity across the economy.

When software becomes easier to build, what remains defensible?

Listening to HiBob’s founder and Wiz’s CPO and co-founder, one theme I took away was the importance of becoming a system of record: the trusted place where an organisation maintains essential business information.

I understand the appeal of that argument. As AI makes software faster and cheaper to produce, a feature becomes easier to imitate. Replacing a system that holds authoritative records, permissions, history and deeply embedded workflows is a much bigger undertaking.

I would hesitate to call it the only moat. Distribution, trust, network effects and differentiated access to data can also matter. But a system of record can combine several advantages: the customer trusts the information, other tools depend on it, and replacing it requires operational change.

Even then, storing data is insufficient. If another product becomes the interface through which customers make decisions and complete their work, the underlying system risks losing some of its influence and pricing power.

My VC takeaway is that founders need to understand where they can build a lasting relationship with the customer. Becoming the authoritative record is one route. Owning a critical workflow, with the context and permissions to complete it reliably, may be another. Either requires a deeper role in the customer’s business than providing a convenient feature.

Juniors may bring an advantage we are overlooking

Another point from the discussion challenged the assumption that AI makes junior hiring less attractive. People entering the workforce have less attachment to how things used to be done. They can learn to work with AI agents from the beginning, without first having to unlearn established habits.

That is a compelling hiring hypothesis. Someone with less experience may be more willing to redesign a task around the tools available today. Experienced colleagues bring the judgement to recognise when the output is wrong, understand the consequences and set the right standards.

The opportunity is to combine those strengths. Give juniors meaningful responsibility, access to AI tools and strong mentorship. Evaluate the quality of their work and how quickly they learn. That still requires investment from managers, especially when AI can produce convincing mistakes.

For Israel, this also matters beyond any individual company. The juniors hired and trained today become the senior engineers, managers and founders of the next decade. An ecosystem that stops developing them weakens its own future.

The organisation has to change too

A final takeaway from the day was that adopting AI requires changes to how organisations work. Buying tools is the easy part. Capturing the value means reconsidering workflows, responsibilities, approvals and the way teams share information.

If an agent can prepare an analysis in minutes but the decision still waits two weeks for approval, the organisation has preserved much of the original bottleneck. The same applies when teams generate more code, content or sales outreach than anyone can usefully review or absorb.

For founders, I think this means measuring the whole process: how long it takes to resolve a customer problem, ship a reliable improvement or complete a transaction. Then redesign the work around the result, with clear accountability and appropriate checks.

This is where Israeli startups have an opening. Small teams can establish these practices early, while larger organisations have more existing systems and habits to change. The advantage will belong to teams that turn faster execution into better products, stronger customer relationships and sustainable economics.

Where I see opportunities for Israeli founders

The following are investment hypotheses prompted by the slides, rather than forecasts established by the research.

1. AI that takes responsibility for valuable work

There is room to build products that complete a defined workflow: resolving a technical problem, processing an operational exception or helping a business convert demand into revenue.

The opportunity becomes more compelling when the product connects to the customer’s systems, handles the difficult cases and produces a result that can be checked. Deep technical knowledge can become a commercial advantage when it makes a costly task materially easier.

Founders should establish who owns the budget, what successful completion means and how much human work remains behind the product. Those answers determine whether a promising demo can become a scalable business.

2. Consumer AI, creative tools and gaming

From Remagine Ventures’ perspective, I am particularly interested in how AI changes the way people spend their time and money.

That includes tools that help creators produce and distribute work, products that help consumers discover and buy, and gaming experiences that become more responsive or personalised. Smaller teams may be able to test products and reach international users earlier.

The difficult part is earning repeated use. Founders need to understand why users return, how they discover the product and whether the economics improve as usage grows. A product that attracts curiosity needs a path to becoming a habit or solving a recurring problem.

3. Turning technical expertise into commercial products

The Innovation Authority’s companies chapter records an increasing share of new company formation in defence and space, alongside quantum and infrastructure. This points to a broader range of technical starting points for company building.

I see potential where specialist expertise can address a recurring commercial need: industrial monitoring, physical security, robotics or the analysis of complex engineering systems.

The test is whether a team can turn that expertise into a repeatable product. Procurement, deployment, regulation and support need to be considered early. A successful pilot becomes much more valuable when the next customer can buy substantially the same product.

4. Helping smaller companies operate globally

The scale-up problem is also a market opportunity. Smaller teams still need to manage customers, coordinate operations, understand performance and expand across borders.

Software that makes these activities easier could help more businesses reach global markets. A strong starting point is a specific customer group with an expensive recurring problem and a clear route to reaching similar buyers.

For founders building these products, distribution deserves attention from day one. Knowing exactly how to reach the next hundred customers can be as valuable as another product feature.

The ecosystem has work to do

The presentation’s framework brings together founders and managers, industry, capital and academia. That is a useful reminder that scaling depends on capabilities spread across the ecosystem.

Universities can connect technical education with commercial practice. Established companies can become customers and training grounds for future operators. Investors can help recruit managers, open customer relationships and support sensible financing decisions.

There are already signs of maturity. The Innovation Authority reports that Israeli companies acquired 81 foreign companies in 2025, while other Israeli companies bought 49% of the Israeli companies sold that year. Acquisitions can give growing businesses access to customers, products and capabilities they would otherwise take years to build.

Yet a strong growth-stage market also needs a healthy pipeline of new companies. The same report records just 301 funding rounds of up to $10 million in 2025, the lowest number in a decade. Bigger headline funding totals can coexist with a tougher environment for emerging teams.

That is one reason I continue to believe in a diverse early-stage funding market. Small funds, angels and specialist investors can support founders whose opportunities are still too early or unfamiliar for larger pools of capital.

Acquisitions also remain an essential part of a healthy ecosystem. They reward founders and employees, return capital and help create the next generation of companies. More founders should have a credible choice to continue building when the market and the business justify it.

Israel’s next chapter will depend on making that choice more achievable: helping founders reach customers, develop leaders and build durable businesses, while making Israel a place where they want to keep investing their talent.

AI may change how many people it takes to build a global company. The quality of the people, institutions and networks around that company will still shape how far it can go.

One last thing

I’m writing this on the evening before the Israeli elections. As someone who lives abroad and invests in Israeli founders, I care deeply about the future they are building and the country my children will know. The conversations today were about growing companies, but they also brought me back to the conditions that make people want to build their lives here: strong democratic institutions, predictable policy, excellent education and confidence in the future. I remain optimistic about Israeli talent and ambition. Tomorrow, we each have a chance to help shape the conditions in which they can flourish. If you’re eligible to vote, please go out and do it. The future we’ve been discussing deserves your voice.

The original version of this newsletter was published on Linkedin to 8,500+ subscribers. Signup for free and be in great company.

Firgun (pronounced “feer-goon”) means taking genuine, unselfish joy in someone else’s success. Every week since March 2020, this newsletter has celebrated the founders, teams and investors moving our community forward. If someone forwarded this to you, you can find previous editions on VC Cafe.

Israel goes to the polls on Tuesday, October 27, and thousands of Israelis living abroad (me included) are travelling home just to vote. Organisers are working to mobilise tens of thousands, with people taking time off work, arranging childcare and spending their own money to make their voices heard. Whatever your politics, that commitment deserves recognition. Democracy depends on people showing up. My hope is that the next government rebuilds trust, protects democratic institutions and gives people the confidence to build their companies, careers and families in Israel as I mentioned to the Jewish News this week.

In the wider tech world, all eyes were on Anthropic as reporting on its draft IPO prospectus revealed $16.3 billion in first-half revenue and an $80.7 billion net loss, largely driven by non-cash financing charges, alongside more than $500 billion in cloud and infrastructure commitments over the next decade. A potential November listing could become a defining test of investor appetite for AI, although its timing and valuation remain unsettled. Anthropic also launched Claude Duet, a proactive personal agent for professionals and small businesses that can prepare briefings, organise meetings and carry on working while users are away. OpenAI, meanwhile, is reportedly nearing a raise of up to $30 billion at a $1.25 trillion pre-money valuation, with Coatue and UAE investors leading, while announcing that its models and Codex workflows are coming to Google Cloud. It also expanded ChatGPT Ads to 26 additional countries, saying more than 100,000 advertisers have used the platform. In China, DeepSeek is wrapping up a roughly $15 billion funding round, double its original target. The mega-rounds continued elsewhere: Tenstorrent raised $1.5 billion at a $6.2 billion valuation, Flapping Airplanes announced a $1 billion Series A at a reported $6 billion valuation, and Legora secured $550 million at an $8.8 billion valuation. The infrastructure race remains relentless: NVIDIA introduced Minerva to help operators monitor AI factories, while Meta put its annual capital expenditure outlook at $135–145 billion. Yet AI spending concerns are delaying other IPOs, making the gap between technological ambition and financial returns harder to ignore. Tesla and SpaceX formed special committees to explore a potential merger, although no deal has been agreed, and Meta was ordered to pay a $25 billion privacy penalty, which it plans to appeal. Finally, critical PHP vulnerabilities prompted urgent patching warnings, while Railway disclosed a possible exposure of customer secrets and urged users to rotate credentials.

QUOTE OF THE WEEK

“Democracy cannot succeed unless those who express their choice are prepared to choose wisely.” — Franklin D. Roosevelt

NEW FUNDING ROUNDS

• Big congrats Amnon Shashua and team doubleAI on coming out of stealth with $220M in funding to pare the cost of compute by a reported 67%!
• Huge congratulations to Raziel Oren, Guy Eshel and team Mave on announcing $34 million in total funding, including a $28 million Series A and previously undisclosed seed round, to buildAI agents that investigate security alerts across companies’ existing tools and help security teams respond more effectively!
• Mazal tov to Benny Meir and team Modelyo on coming out of stealth with $17 million in seed funding to develop confidential cloud technology to help organisations use AI and cloud infrastructure while retaining control over sensitive data, including while it is being processed!
• Kol hakavod to Alon Abelson and team High Lander on securing an additional $16 million, bringing total funding to $30 million to build a drone fleet and low-altitude air traffic software!
• Way to go Dvir Zeltzer, Eli David, Ph.D. and team Two Delta on reported seed funding of tens of millions of dollars to develop technology that accelerates AI inference and improves computing efficiency!

EXITS

• Huge congratulations to Maayan Salom, Or Avidov and team Elementary on your acquisition by Boomi, in a deal reported at approximately $40–45 million in cash and stock to help companies monitor the quality and reliability of the data behind business workflows and AI agents!
• Mazal tov to Elon Gliksberg and team HexAI on your acquisition by Hex-Rays, in a deal reported at $30–35 million! Founded just last year, the six-person startup built an operating system for AI agents conducting security research without raising outside capital.
• Kudos Yoni Cohen and Eylon Zefania on the acquisition of Vizby AI by Swap for an undisclosed sum and in less than a year for your GEO platform for commerce!

NEW FUNDS

• Congratulations to Abe Finkelstein, Asaf Horesh and team Vintage Investment Partners on raising $920 million across a $580 million fund of funds and a $340 million secondary fund!
• Is it too early to congratulate Esther Barak Landes and Adi Ostry Matalon on the launch of ALMA Ventures, a new $40M early stage fund focused on mental health and longevity in partnership with Sheba Medical Center, Tel Hashomer?

LINKS FOR YOUR BROWSER

VC CAFE / REMAGINE VENTURES

• Venture Capital in the Age of Consensus – My take on capital concentration, extraordinary pre-product valuations and why smaller funds still have an important role in backing overlooked opportunities.
• Who Pays for the AI Boom? – AI can deliver enormous value while investors struggle to earn a return. A look at the assumptions behind the spending.
• The Rise of the Inference Economy – The companies turning models into usable services, and why their economics differ from simply renting out GPUs.
• I launched Israel VC Jobs board to help people find jobs at venture-backed Israeli startups. It already links to 2083 open jobs. Check it out here and find out more about the why IsraelVC Jobs here.

ISRAEL

• The Q3 2026 tech review report by IVC Data and Insights
• Investors turn to Israel’s battle-tested defense-tech start-ups as funding hits record $40.8b – Investments are pouring into military and dual-use technologies at unprecedented levels due to rising military spending and demand for autonomous systems and deployable battlefield technologies.
• IDF establishes Space Division as Israel eyes warfare beyond Earth – The new unit will bring together military projects involving orbital sensors and space-warfare technologies. The move formalises the IDF’s growing focus on space, while leaving open how quickly its longer-term ambitions for military operations beyond Earth can be realised.
• Nvidia unveils Israel-developed AI agent protection architecture – CEO Jensen Huang unveiled a new security architecture at the GTC conference in Berlin, which has an Israeli development core.

GLOBAL

• The View from Here – Larry David’s modest proposals for the Democratic Socialists of America. Witty and light view on the ‘bagel people’ :-)
• On risk – a candid view by Jake Gibson on how warped venture is today, where VCs and founders are optimising for what the next round VCs are looking for. Micah Rosenbloom warns that the same trend didn’t end well in the dot com bubble.
• Also on the state of ventures (and in the spirit of this will not end well) Bloomberg reports that investors are paying different prices for the same startup, often at the same round. AI fomo is driving investors to do silly things to get into ‘hot’ deals.

MEDIA OF THE WEEK

The Q3 2026 IVC report highlights (source) – back to market growth, but concentration continues

The inference ecosystem landscape – My updated map of the companies helping turn AI models into useful, deployable services.

What does good growth look like in 2026 – the Pacesetter index by ICONIQ (source)

Greenfield Disruptors 60 by Greenfield Partners (see the full ranking)

Where do AI deployment firms focus by Elaia (source)

The periodic table of business functions by Idealab.ai – a collection of 1.2 million free ideas by Bill Gross (source)

Made me chuckle ;-)

That’s all for this week. Behind every announcement are people who spent months or years building, recruiting, selling and working through setbacks. Give them a little firgun. A thoughtful message, an introduction or a public word of encouragement can go a long way.

Keep on creating!

Eze Vidra

Harry Stebbings called it out:

The squeeze also extends to the middle of the market. According to Murph Capital’s Q3 fundraising review, citing PitchBook data, funds between $100 million and $500 million accounted for roughly 17% of US fund closes before 2021. Through Q3 2026, their share had fallen to 10%, the lowest on record. For founders, a healthy funding ecosystem needs depth: investors with different fund sizes, ownership targets and definitions of a meaningful outcome.

Each of these developments has an explanation. Taken together, they raise a more uncomfortable question: how much of venture is still about discovering potential, and how much has become about buying access to what everyone already agrees will matter?

As a pre-seed investor, I find parts of this market frustrating. But I’m also encouraged by what founders can now accomplish. Capital is becoming more concentrated just as the ability to build is becoming more accessible. That tension creates opportunities worth paying attention to.

The age of consensus capital

Nico Wittenborn captured the mood with the phrase “the age of consensus capital”. The underlying figures are striking. According to the Q1 2026 PitchBook–NVCA Venture Monitor, five venture firms captured 73.1% of US venture fundraising during the quarter. Five deals accounted for 73.2% of US venture investment, or $195.6 billion. These are quarterly US figures, but they illustrate how narrow the flow of money has become.

Concentration at both ends matters. When fewer firms control more capital, their economics increasingly determine which opportunities get funded. A business capable of producing an excellent outcome for a small fund may barely register for a multibillion-dollar vehicle. The company hasn’t become less useful to its customers. It has become less relevant to the investor’s fund size.

There are rational reasons for LPs to favour established managers and for investors to back apparent category leaders. Some of those leaders will build extraordinary businesses. But an industry can make individually defensible decisions and collectively become less adventurous. If everyone needs exposure to the same companies, independent judgement risks giving way to the fear of being absent.

Venture has always depended on a small number of exceptional outcomes. The harder question is whether we are becoming too certain, too early, about where those outcomes will come from.

When traction struggles to compete with possibility

Many software founders find themselves in a strange position. Revenue and growth provide evidence that a business is solving a problem. Yet they can feel insufficient when an investor’s first question is whether AI will make the entire category disappear.

That question deserves to be asked. Existing revenue does not guarantee future relevance. AI can change pricing, weaken differentiation and shift value away from established software categories. Founders need a convincing answer about why customers will continue to choose them.

But scepticism should be applied consistently. We should scrutinise the durability of an existing business and the assumptions behind one that has yet to prove itself. Sometimes the market seems to demand certainty from the former while granting almost unlimited possibility to the latter. A company with customers has measurable imperfections. A company with a compelling vision can still be imagined as perfect.

Consider Business Insider’s reporting on Discovery Loop, founded by former Google chief scientist Jeff Dean and other prominent researchers. In September, it was reportedly seeking a valuation of around $50 billion, before releasing a product. That was a fundraising target, rather than a confirmed completed round.

There is an important caveat here. These valuations belong to a rarefied corner of venture: the new AI research labs, or “neolabs”, founded by people with exceptional research credentials. They are not representative of what an ordinary AI startup can raise. Investors are underwriting scarce talent, technical ambition and the possibility of a major scientific breakthrough.

That deserves serious consideration. Ambitious research often requires funding long before commercial validation. Pre-product investing is fundamental to venture, and as a pre-seed investor I would hardly argue otherwise.

The question is how much future success is already embedded in the price. At a $50 billion entry valuation, a tenfold increase implies a $500 billion company, before subsequent dilution. That doesn’t make the investment wrong. It does make the expectations extraordinary.

What happened to partnership?

The strain extends beyond valuations. TechCrunch reported that Benchmark’s partners acknowledge growing portfolio conflicts as AI companies expand into overlapping markets.

Two companies can become competitors through changes neither originally anticipated. That is different from deliberately financing a direct rival after gaining access to a founder’s plans. The distinction matters to anyone choosing an investor.

From the founder’s perspective, the questions are practical. Can I discuss a product roadmap openly? Can I explain why we lost a customer? Can I share an acquisition opportunity without wondering whose interests take priority? Confidentiality is essential, but trust also depends on attention, incentives and judgement.

Investors should be explicit about the relationship they offer. If the promise is close partnership, the obligations should reflect that. Founders deserve clarity before accepting the cheque.

A related issue sits underneath the pro-rata debate covered by Axios, following the dispute between Valar Atomics and Day One Ventures.

Without taking a position on that particular case, I think the credibility of an early agreement matters. Founders and investors should be able to negotiate changes as a company develops. Those conversations become harder to trust if bargaining power is the only principle that survives success.

Investors also need to earn their place through how they behave. A signed agreement cannot substitute for being useful, responsive and fair. Equally, a founder’s success should not automatically turn an early partner’s negotiated economics into an inconvenience. The ecosystem needs people willing to commit before the outcome is obvious, and relationships that remain workable afterwards.

Why I’m still optimistic at pre-seed

There is still an enormous amount that technology can improve. Small businesses spend too much time on administration. Healthcare professionals wrestle with fragmented workflows. Education struggles to adapt to individual needs. Creative work remains constrained by production costs. Consumers navigate services that are unnecessarily difficult to understand and use.

None of these problems becomes less important because most of the capital is flowing somewhere else.

What encourages me at pre-seed is how much further a small team can get before needing a large organisation. AI tools can help founders write code, explore designs, analyse information and test ideas with customers. An early cheque can potentially fund more learning, with fewer resources committed before the team understands what people actually need.

That does not make company building easy. Lower barriers to building also mean more competition. Distribution, customer trust, retention and reliable execution remain difficult. A working demo can arrive quickly; a business that customers depend on still has to be earned.

But the ability to run more experiments with less capital is meaningful. A team can pursue a neglected workflow, learn from a narrow customer group and expand from a position of understanding. It does not need to begin with a consensus view of an enormous market to discover one.

At Remagine Ventures, backing Israeli founders at pre-seed, that is the work I remain excited about: meeting people before the story is polished, understanding something others have overlooked, and helping turn an early insight into a business.

Small funds have to be disciplined about their own economics. Being overlooked is not proof of investment potential. But we can give serious attention to opportunities that are initially too small, too unfamiliar or too early for much larger investors.

Venture needs room for different fund sizes, different convictions and different paths to success. We should want ambitious research labs to succeed, while making space for founders whose strongest credential is how well they understand a customer’s problem.

I remain optimistic because the supply of important problems is enormous, and the tools available to solve them are improving. Capital may be clustering around the same names, but the next generation of founders does not have to wait for consensus to start building. Our job at pre-seed is to recognise their potential before consensus arrives.

That distinction is becoming harder to ignore. Anthropic’s anticipated IPO, with reported valuation ambitions reaching $2 trillion, the semiconductor selloff in Asia and rising concern about neocloud debt are testing different assumptions behind the AI boom. Together, they raise a question that reaches well beyond public markets: how much of the ecosystem can sustain itself once customer spending has to carry more of the weight?

Sophie Shulman’s analysis in CTech connects these three pressures. For venture investors, the implications extend to where capital goes, how valuations are set, what counts as traction and how diversified our portfolios really are.

Extraordinary growth can still disappoint

Anthropic provides perhaps the clearest illustration. According to reporting by the Financial Times, the company told prospective investors that quarterly revenue reached $16.8 billion in Q3, up from $11.5 billion in Q2, alongside profitability on an adjusted basis. Or as Henry Blodget reported, while the growth of Anthropic has truly been unprecedented, it might be stalling. The company added more new revenue in Q2 2026 than Q3.

Q1: $4.8 billion
Q2: $11.6 billion (+$6.8B and +141%, sequentially)
Q3: $16.8 billion (+$5.2B and +45%, sequentially)

That is approximately 46% growth in three months, but media reports have mentioned that the expected price tag of roughly $2 trillion, is more likely to be closer to $1.5 trillion – still the biggest IPO of all time, but nevertheless a big haircut. These remain reported discussions, rather than a final IPO price. In the meanwhile, OpenAI is looking to raise $30 billion at $1.4 Trillion valuation ahead of said IPO, perhaps also with the goal of creating a clear pricing milestone at this point in time.

A business can grow exceptionally quickly and still fall short of the assumptions embedded in its valuation. That is why I find “is Anthropic too big to fail?” less useful than asking what has to happen for investors buying at these prices to earn an attractive return.

The accounting deserves scrutiny too. “Adjusted profitability” is not the same as free cash flow. Investors need to understand which costs are excluded, how revenue is recognised and what commitments sit behind it. Stock-based compensation, where excluded, still has a cost to shareholders through dilution.

At the same time, today’s economics are an imperfect guide to the eventual business. Cheaper inference could improve margins or be passed to customers through competition. Agents could generate much more usage while increasing the cost of serving each customer. Training requirements could change again.

Today’s numbers matter. They need to be examined alongside scenarios for how the business evolves, rather than extrapolated indefinitely.

The funding model is becoming part of the story

There is an important counterweight to the bubble debate: products such as coding agents give customers concrete reasons to use and pay for AI. Commercial adoption and financial fragility can develop simultaneously.

The question is whether the infrastructure supporting that adoption earns enough, quickly enough, to justify its cost.

Semiconductor investors are already reassessing expectations. CTech reported that the Philadelphia Semiconductor Index had fallen 23% from its peak, while Reuters described sharp selling in Asian chipmakers, including South Korean stocks and Japan’s Kioxia.

The pressure is particularly visible in neoclouds, the specialist providers financing and renting out GPU capacity. CTech reported roughly $50 billion of debt at CoreWeave, with yields on some bonds reaching around 15%. That market yield is not the interest rate on all existing borrowing, but it signals how investors are reassessing risk.

As expansion becomes more dependent on external capital, lenders and public-market investors gain greater influence over its pace. A project can have interested customers and still struggle to attract financing on acceptable terms.

I would be careful about claiming that every hyperscaler has exhausted its capacity to self-fund. Their balance sheets and cash generation differ. But the broader shift matters: financing conditions increasingly help determine how much infrastructure gets built, by whom and when.

When your investor is also your supplier

The circular financing debate sits at the centre of this shift. Nvidia is a major strategic investor in neoclouds, while also supplying the technology on which their businesses depend.

The documented relationships are substantial:

This graphic by Bloomberg, published in October 2025, illustrates how tangled the financial investments in AI are:

The mechanism is straightforward. A supplier invests in a customer. The customer raises more capital and buys the supplier’s technology. Those purchases support the supplier’s revenue and expand the market for its products.

There are legitimate reasons for these arrangements. Infrastructure requires upfront investment, engineering partnerships can improve performance, and suppliers may understand the opportunity better than generalist investors. Nvidia argues that financing enables infrastructure for real demand, with financing partners assessing utilisation, cash flow and residual value.

Supplier investment alone is not evidence of artificial revenue. But it changes the diligence question: how much activity is ultimately supported by customers earning or saving money outside that financing network?

Imagine a startup raises $100 million and spends $40 million on compute. Its provider records revenue and purchases equipment. These can all be legitimate transactions, but they are not independent proof of end-customer demand. Some of the activity is being financed by the same original pool of capital.

GPUs are valuable collateral, but their value can fall

The argument that GPU-backed lending is exceptionally safe rests partly on the usefulness and scarcity of the hardware. Yet collateral is only worth what another buyer will pay when a lender needs to sell it.

If compute rental prices fall, the income generated by a GPU can decline. Its resale value may fall too. The borrower’s cash flow and the lender’s collateral can therefore weaken together, precisely when protection is most needed.

For now, the demand for compute is still outpacing the supply, causing GPU prices to go up. As reported by A16Z:

Rental rates (and thereby residual values for GPUs) are supposed to decrease over time, but that’s not really what’s been happening. As intelligence gets cheaper, demand for compute is only rising, causing GPU pricing to climb upwards for the latest chips (and stay pretty for the older ones). Even the A100 is pricing at-or-above what it was at the beginning of the year.

The chips do not need to become obsolete for this to hurt. They can remain productive while being worth considerably less than the financing originally assumed.

Long-term customer contracts help, but their protection depends on counterparty strength, cancellation rights, delivery conditions and margins. They also do not remove construction delays, power constraints or refinancing risk.

In Reuters’ October 19 reporting, S&P Global warned that interconnected financing structures could leave major technology companies with more AI-related liabilities than were currently visible. It did not quantify the additional exposure or see an immediate need to change ratings.

Parts of this market require infrastructure and credit underwriting alongside venture judgement. Funding an uncertain product and financing a fleet of depreciating machines demand different questions.

Could Wall Street prevent the glut it fears?

There is a potentially self-correcting mechanism here. If lenders fear excess capacity and stop financing new projects, fewer projects get built. That could reduce the very oversupply they worry about.

But timing matters. Capacity already under construction may still come online. Existing debt still needs servicing. Demand could weaken faster than supply adjusts, and a company can run out of cash before the market reaches a healthier balance.

Financing discipline could therefore improve the economics of future infrastructure while inflicting losses on investors who funded the current buildout. A sustainable long-term market does not guarantee that every participant survives the transition.

This also means a slowdown need not affect everyone equally. Operators with strong customers, manageable obligations and operational capacity could fare very differently from those still financing construction or relying on repeated fundraising.

What the boom is doing to venture capital

The effects reach far beyond infrastructure. According to the Q3 2026 PitchBook–NVCA Venture Monitor, AI accounted for 82.7% of US venture deal value through September. Funds of at least $500 million captured 78.1% of fundraising capital while representing just 6% of funds closed.

These are shares of dollars, not shares of startups. They help explain how record investment can coexist with difficult fundraising for smaller managers and companies outside the most sought-after deals.

For early-stage investors, one danger is importing the economics of enormous AI rounds into every seed investment. A frontier lab, a GPU cloud and an application startup may all carry the AI label, but their capital requirements, margins and routes to an exit differ fundamentally.

Strategic investors also have a different return equation. A chipmaker can benefit from equity appreciation and hardware sales. A cloud provider can benefit from equity appreciation and compute consumption. A venture fund primarily needs its ownership to turn into cash proceeds. The same investment price may make sense for one investor and much less sense for another.

Concentration can also hide behind apparent diversification. S&P Global Market Intelligence has highlighted overlapping exposure to AI companies across different investment managers.

I would take that concern further. An LP could own a model developer through venture, a neocloud through growth equity, data centres through infrastructure and related loans through private credit. Different fund labels may still conceal dependence on the same customers and spending assumptions.

Meanwhile, higher private valuations do not automatically produce liquidity. A funding round can improve reported returns without distributing cash to LPs. That distinction matters for the firms trying to raise their next fund and the founders depending on them.

The opportunity for early-stage founders

As an investor backing Israeli founders at pre-seed through Remagine Ventures, I remain excited about making these capabilities useful. I also believe a healthy ecosystem needs investors willing to back non-obvious companies before they become consensus bets.

For those companies, revenue quality deserves as much attention as revenue growth. Revenue from venture-backed customers is real, but customers dependent on frequent fundraising may cut spending together. A startup can appear diversified by customer count while remaining exposed to a single financing cycle.

Margins deserve the same scrutiny. Cloud credits and cheap pilots can obscure the cost of production workloads. Retries, human review, support and heavier usage all count. The useful measure is the cost of reliably delivering the result the customer wants.

For consumer AI, gaming and creative tools, that means understanding whether novelty turns into retention and willingness to pay. More content generated or more tokens consumed can signal engagement, but neither establishes a durable business on its own.

The questions I would keep returning to are:

• Does the customer renew because the product delivers a meaningful benefit?
• Do margins hold after credits expire and real operating costs are included?
• Does the business become stronger as models improve and inference gets cheaper?
• Is differentiation accumulating through distribution, workflow integration, proprietary data or trust?
• Can the next milestone be reached without assuming another unusually generous funding round?

The AI boom is creating extraordinary capabilities and a financing network in which suppliers, customers, investors and lenders increasingly overlap. Understanding both is becoming essential to understanding the opportunity.

I remain optimistic about what founders can build. The discipline is to keep asking where the cash ultimately comes from, who retains the margin and what must happen for shareholders to earn a return. The dot-com era showed how profoundly a technology can succeed while individual investments fail. We should be ambitious enough to back the opportunity and clear-eyed enough to remember the difference.

That distinction is becoming harder to ignore. Fireworks has disclosed more than $1 billion in annualised revenue. Modal crossed the same threshold this month. Together AI has reached a similar scale according to external estimates, while fal has reportedly reached an $800 million revenue pace serving generative media.

Behind those numbers is a shift in where value can accumulate. The race to train frontier models continues, but a growing group of companies is building businesses around deploying them. Some model developers are joining that market themselves, including by serving open-weight models developed in China.

Nathan Benaich captures the tension in the State of AI Report 2026:

“either you die trying to get to the frontier, or you live long enough to serve inference.”

For founders and investors, the question is what makes that second business defensible. Following my post on the rise of Neoclouds (and the timely headlines about CoreWeave losing a bunch of value given its high debt) I wanted to dive into the new category of startups selling Inference.

What inference providers actually sell

Training creates a model. Inference runs it: generating an answer, producing an image, analysing a document or deciding an agent’s next action. Every useful application needs inference, whether its developer builds the underlying model or accesses one through an API.

This is related to the neocloud market I have been exploring, but the customer buys something different. A neocloud primarily sells computing capacity: GPUs, networking and the infrastructure needed to run workloads. An inference platform packages that capacity into a service for running models. Its promise is that developers can ship an application without assembling and operating the entire serving stack.

The boundaries overlap. Together sells GPU capacity as well as model APIs. Modal supports inference, training and agent execution. Cerebras and SambaNova combine specialised chips with systems and services. Calling all of them “GPU clouds” misses the differences that determine their economics.

The companies turning model usage into revenue

The financing and revenue figures show how substantial this market has become. They also require care: annualised usage revenue, bookings and recognised quarterly revenue measure different things.

Sources: company announcements from Fireworks, Modal, Together, fal, Baseten, DeepInfra, Groq and SambaNova; Cerebras’s SEC-filed results; Sacra on Together and Baseten; and The Information on fal. Figures have different observation dates and are not a revenue ranking.

Two details are particularly revealing. Fireworks says more than 95% of its daily tokens come from models specialised on customer data and specific jobs. Its opportunity extends into helping customers adapt models to their own work. Modal’s October announcement puts its sandbox product alone at a $300 million run rate. Agents need environments in which to execute code and use tools, creating infrastructure demand beyond generating tokens.

These businesses are expanding towards the surrounding workflow. That matters because an API offering the same model as everyone else faces persistent price competition.

When model builders become model hosts

The most interesting development is that model research and inference are becoming complementary businesses, including for companies originally associated with building their own models.

Reka is a clear example. Its 2024 technical report described multimodal models trained from scratch, with its largest model positioned against leading frontier systems. Its commercial offering now also includes Reka Cloud: serverless inference, dedicated GPUs and model customisation. The models displayed include DeepSeek, Qwen and GLM. A company that developed its own intelligence is also selling the infrastructure to run intelligence developed elsewhere.

That is an expansion, rather than evidence that Reka has abandoned research. Its current business still includes a model research lab. The important shift is that monetisation need not depend exclusively on whether its own next model wins.

Zyphra makes the connection explicit. Alongside developing its own foundation models, it launched an inference cloud on AMD infrastructure supplied by TensorWave. Its launch catalogue included DeepSeek, Kimi and GLM. Founder Krithik Puthalath described the cloud as “the natural extension of our research.”

Here, research expertise becomes a commercial capability: understanding models well enough to run them efficiently on a particular hardware stack. Zyphra’s announcement also identifies the infrastructure partner. Offering an inference cloud does not establish that the provider built or owns the underlying data centre.

01.AI represents a more explicit change in direction. Kai-Fu Lee’s company developed the Yi model family before shifting towards enterprise deployment in what he called the DeepSeek age. The Financial Times reported its pivot, while Reuters covered its Wanzhi enterprise platform. The lesson is that capable open models let a company commercialise deployment and domain expertise without funding every successive pre-training race itself.

These examples support a shift from exclusive reliance on proprietary models towards deployment businesses. They do not all support the stronger claim that former frontier labs have stopped research and become owners of giant data centres.

Poolside illustrates the separate move into physical infrastructure. Its Project Horizon announcement with CoreWeave described a planned 2GW Texas campus. But the FT reported in April that the partnership had ended, and Poolside continues releasing its own models. It is an example of model development intersecting with infrastructure ambition, rather than proof of an operating Chinese-model hosting pivot.

Chinese models are becoming inputs to Western AI businesses

There is an irony here. The public debate often frames AI as a contest between American and Chinese model companies. The commercial stack is more interconnected: a Chinese-developed model can run on American-designed chips, through a Western inference provider, inside an application sold to a global enterprise.

Open weights make that separation possible, subject to each model’s licence. They give other companies the ability to deploy and adapt the model, so the organisation that creates it does not necessarily collect every subsequent inference dollar.

The origin of a model and the location of its deployment are also different questions. Running downloaded weights with a chosen operator can keep processing within that deployment; it is different from calling the original developer’s hosted API. Customers still need to understand who operates the service, where data is processed and what controls apply.

For inference providers, better open models expand the range of workloads they can compete for. For application founders, the opportunity is to combine those models with proprietary data, evaluation and distribution. Neither requires winning a general-purpose benchmark.

Cheaper intelligence can create a bigger bill

The State of AI report highlights Epoch AI’s research, which estimates that the cost of achieving a fixed score across five benchmarks has fallen by about 47% per quarter since 2023.

That does not tell us whether total spending falls. A cheaper operation becomes economical in more settings, and an agent may perform many operations to finish one assignment. A business that once generated a draft can now explore alternatives, call tools, check results and retry. The relevant comparison becomes the cost of a completed task.

A cheaper model can therefore produce a more expensive workflow if it takes more attempts or requires more human review. Equally, an expensive model may be worth using for one difficult planning step while cheaper models handle repetitive execution. Founders need evaluations that reflect their actual product, rather than a procurement decision based only on token prices.

Where the moat might be

My view is that inference can support large businesses, but their durability will depend on what customers would lose by switching.

Research expertise can improve performance and cost. Deployment software can remove weeks of engineering work. Enterprise controls can make a service usable for a customer who could not otherwise adopt it. Specialisation can align the platform with a particular workload, such as generating video, handling voice interactions or running long-lived agents.

The test is whether those advantages persist as models, chips and open-source serving software improve. A provider whose value is mostly temporary access to scarce GPUs has a different business from one embedded in how customers train, evaluate, deploy and monitor their models.

Revenue alone will not answer that question. I would want to understand gross profit after compute costs, customer concentration, infrastructure commitments and retention when a major customer changes its preferred model. Annualising a strong usage month cannot establish the quality of the underlying business.

Physical infrastructure adds another set of obligations. Power contracts, financing announcements and planned campuses precede usable capacity, sometimes by years. As the State of AI infrastructure analysis emphasises, contracted power and operating capacity are different measures. Buying into the inference opportunity does not remove construction, utilisation or financing risk.

For the Israeli founders we meet at Remagine Ventures, the practical implication is encouraging. You can build a valuable AI product using models that keep improving outside your company. The work is to own the customer relationship, accumulate useful feedback and deliver an outcome at economics that improve with scale.

That applies whether you are building creative automation, a consumer application or enterprise infrastructure. Start with the model that lets you learn fastest. As usage grows, earn the right to take more control of the stack through better data, better evaluations and a measurable benefit to the customer.

The frontier race will continue to attract extraordinary capital. The inference economy gives founders another route to a large business: make increasingly capable intelligence work reliably for customers, and retain enough value from doing so.

The original version of the #FIRGUN Newsletter was published on Linkedin and emailed to 8,500+ subscribers. Sign up to be in good company :-)

Firgun (pronounced “feer-goon”) means taking genuine, unselfish joy in someone else’s success. Every week since March 2020, this newsletter has celebrated the founders, teams and investors moving our community forward. If someone forwarded this to you, you can find previous editions on VC Cafe.

With Israel heading to the polls on October 27, there’s a lot at stake for the country and its future. For the tech ecosystem, the next government’s decisions will help shape investor confidence, the ability to attract and retain talent, and Israel’s relationships with its global partners. My hope is for leadership that rebuilds trust, brings people together and gives founders the stability to keep building. The ingenuity and determination are there. They deserve the conditions to flourish.

In the wider tech world, flying cars finally had their moment: Tesla demonstrated its next-generation Roadster, using SpaceX-developed thrusters to perform airborne manoeuvres, although production is still targeted for 2028. Meta released Muse Spark Open, making the weights of its new coding and agentic model available to developers, with support for running it locally on two Apple M5 Ultra Mac Studios. That followed its launch of Muse Wonder and Wonder Full, while Google opened Gemini 4 Argon to developers for complex coding and enterprise workflows. The funding rounds were equally eye-catching: Ramp raised $1.85 billion (!) at a $60 billion valuation, and Labelbox announced more than $100 million in Series E financing, including equity and a credit facility, to expand its work in reinforcement learning and enterprise agents. Anthropic told the press that they’ve become profitable for the first time this month with $16 billion in profit. The LLM published red-team research showing that small open-weight models could in principle self-replicate across cheap IoT devices (tested on 15 Android TV boxes), a warning that it expects to become practical by next year, and said it has cut extracted reasoning tokens 90-99% with new distillation defences. OpenAI had a stranger week: the FBI arrested a former product-team employee accused of planting a recording device in an executive conference room at its San Francisco office, the company reportedly cancelled a GPT-6.1 Astra model and the press questions a ‘missing’ $20 billion in reported revenue. Microsoft unveiled its Maia 300 chip (eight HBM4 stacks, 288GB), with multi-gigawatt deployment starting in 2027, and Character.ai filed for Chapter 11. Out of China, DeepSeek is reportedly nearing a 100 billion yuan (about $14B) round at a 500 billion yuan valuation, double its original target. Chip stocks wobbled on an IMF warning about an AI investment collapse, with SK hynix down 10% in premarket, while IDC put NVIDIA‘s China data-center AI chip share at 34% versus Huawei‘s 28%. Beyond the commercial race, CERN and Mistral teamed up to develop Flare, an AI model for particle physics. Apple launched HomeView, a new tablet for the family.

QUOTE OF THE WEEK

“Too bad that all the people who know how to run the country are busy driving taxicabs and cutting hair.”

George Burns

NEW FUNDING ROUNDS

• Huge congratulations to Shahar Bahat ?? , Tal Ofer Regev ??, and team Pluto Security on emerging from stealth with $42 million in total funding, including approximately $34 million in its latest financing to help enterprises secure the AI tools, applications and automations employees are adopting and building!
• Mazal tov to Guy Weiss, Jonatan Bien and team Zenzap on unveiling $44 million in funding, comprising an $11 million seed round and a $33 million Series A to bring employees and AI agents together in a shared messaging workspace!
• Way to go Yaniv E. and team Undecimal Bio on an $18 million seed round to deliver RNA medicines to specific cells beyond the liver, using experimental data and AI to identify promising delivery candidates.
• Kol hakavod to Thai (Elgrabli) Sade, Avi Keren and team BloomX on a $13 million Series A to combine robotics and AI to help growers pollinate crops, including blueberries and avocados!
• Kudos Jared Zelman, Raphael Assor and team Othello on coming out of stealth with $8M in seed funding to scale their AI-powered real-time sales coaching platform!
• Great stuff Gilad Salinger, Dror Wolmer and team Uvi on emerging from stealth with $7.5 million in seed funding to help AI agents access and retain company context!

EXITS

• Huge congratulations to Yaki Faitelson and the Varonis team on signing an agreement to be acquired by Proofpoint for approximately $6.3 billion in cash!
• Mazal tov to Gil Peleg, Dan Shprung and team Geniez AI (Acquired by Rocket Software) on their acquisition by Rocket Software, and a reportedly value of approximately $100 million to build AI agents for managing and modernising mainframe systems!
• Congratulations to Tom Laufer and team Loops on the announced acquisition by Salesforce, with the team and technology set to join Tableau to help businesses understand why their metrics changed!
• Congratulations to Uri Haramati and the Torii team on signing an agreement to be acquired by Freshworks for an undisclosed sum to help companies understand which software they use, who uses it and what it costs, increasingly including AI consumption. Financial terms were not disclosed.

NEW FUNDS

• Congratulations Ron Alpar on the launch of Aliyah Ventures, focused on early stage immigrant founders building technology companies in Israel! Best of luck!

LINKS FOR YOUR BROWSER

VC CAFE / REMAGINE VENTURES

• The neocloud boom and the race to outlive the GPUs – My look at the specialist GPU cloud providers powering the AI boom, and the financing and hardware economics behind their growth. There were 209 neoclouds tracked a year ago compared to 323 today.
• The AI opportunity hiding between capability and adoption – Why turning model capabilities into dependable workflows remains a substantial opportunity for founders.
• Get on Remagine’s radar – Israeli founders building at pre-seed, including in stealth: we’d love to hear what you’re working on.

ISRAEL

• Amazon is in talks to acquire Decart for $7 billion – after the Nvida talks fell through, there could yet be a massive exit for the Israeli AI startup
• LinkedIn’s Top Startups list for Israel 2026 – check out the list and methodology
• Tech market changes squeeze out Israeli VC – I disagree with some of the negativity of this @globes article. I still believe that small firms like Remagine Ventures can back the non-obvious very early stage bets and that more diversity is needed in the funding sources

GLOBAL

• KPMG’s Q3 venture update – Global venture investment reached $170.9 billion across 8,341 deals in Q3, with AI megadeals continuing to concentrate capital.
• What is bigger than AI? – 11% of US adults now on GLP-1s (was 3% in 2024); more Americans pay for GLP-1s than pay for AI (a16z).
• The 100 most promising European startups accoding to Sifted and N47

MEDIA OF THE WEEK

Startups worldwide raised $184B in VC in Q3 2026, up 45% on the same quarter last year ($127B). It was the four highest quarter for VC funding on record, with three of the top four coming in the last three quarters. Source: Dealroom.co

We are in the age of consensus capital:

1- Almost 75% of all LP $ raised by 5 funds

2- Almost 75% of all VC $ invested in 5 companies (source)

The neocloud boom and the race to outlive GPUs… (source)

The evolving AI agents by category map by MightyBot (source)

Discussions (and fights) on X this week put to question some of venture capital’s core practices (source)

One of the coolest product launches in recent memory – the new flying Tesla Roadster

Watch the new Apple HomeView commercial. Would you buy one or is this a potential flop?

That’s all for this week. There’s a lot to celebrate in our “small” community! If someone you know has had a win, give them a little firgun. Building is hard, and recognition goes a long way.

Keep on creating.