In my last post, we found that ARR doesn’t mean what it used to. It turns out the metric one layer down has the same problem. So I went back to the VCs who brought up gross margins last time and asked what they look at beneath the headline number. I also asked some founders how they calculate the number in practice.

Shaun is saying that companies can report the exact same gross margin percentage and be describing two entirely different businesses, one still figuring out its pricing, the other still figuring out its balance sheet. Gross margin used to be one of the few numbers on which everyone agreed on how to read. In AI, it’s become a proxy fight over what’s sitting inside COGS, and the investors we talked to have learned to ask a different follow-up question before they’ll believe it.

Token consumers vs. token producers

For applications, the token consumers, Shaun expects margin to be weak early and to “catch up meaningfully“ as the product matures and pricing power builds.

But he also wants to know what sits outside the number. He asks to see “the breakdown of compute costs that may be embedded below gross margin in OpEx“ to understand whether COGS is fully loaded or whether meaningful compute costs sit below the gross margin line.

For token producers, including neoclouds, Shaun puts much less weight on gross margin. “Gross margins are certainly helpful, but we also place higher emphasis on understanding unit-level economics, especially vis-à-vis interest / debt schedules and assumed depreciation rates.”

For capital-intensive infrastructure, gross margin only tells part of the story. Debt schedules and depreciation assumptions can matter just as much.

The labor line nobody agrees on

If compute is the first complication, forward-deployed engineers are the second, and where they sit is a judgment call somebody in finance makes every month.

The distinction matters because the latter starts to look less like software and more like a services business.

A founder at a growth-stage enterprise AI company walked us through what that allocation looks like from the inside. Compute costs get tagged by production versus internal development, production versus pilot accounts, and spend type. Some client costs are directly assigned while others are allocated. FDE and delivery costs are tracked the same way, “at the resource level by the client(s) they support, with time tracking as needed.“ A technical resource dedicated to a client is fully included in Cost of Revenue, while other roles are split based on the percentage of their time that is client-facing.

This founder thinks this calls for “a singular owner to fully manage cloud spend across the business,“ given how much it shapes the margin line and client experience.

Some founders think gross margin works just fine

A founder of a gaming platform pointed out that the cost itself isn’t always known in real time: “It can become very complicated in some scenarios because we don’t even know the exact liability at inference time for certain models.”

Some providers offer rebates once spending crosses certain thresholds, so before the actual bill is issued, the company deducts a “best-guess/worst-case scenario” cost from user balances to keep the number honest in real time.

That uncertainty shows up directly in the margin. Pricing starts at a 50% margin by doubling token costs, but bulk discounts to high-volume customers pull the realized average down to 33%. The pricing model and reported margin can diverge before you even get to accounting choices.

This founder pushed back on the entire angle we’d been hearing: “I honestly think headline GM is a good indicator of customer willingness to pay and your moat beyond a ‘token reseller.’“

A growth-stage fintech founder made a related but different case. “We include product compute and the portion of AI costs directly tied to client delivery in COGS, so those costs are reflected in our reported gross margin,“ he told us.

His company’s COGS, however, are still mostly human labor. In his telling, AI isn’t adding a new cost line. It’s making the labor cost that was already there go further. “The more useful question for investors is not simply how much a company spends on AI, but whether that spend increases delivery capacity and improves the underlying economics.“

Shaun and Vivek want to reconstruct the headline before interpreting it. The founders push back on the idea that gross margin itself is broken. If the costs are classified consistently, they argue, the number can still tell you something important about willingness to pay and delivery economics.

My Take

I still care about gross margin. I wouldn’t underwrite an AI company off the headline number without reconstructing it first.

I’m also skeptical of the argument that low gross margins today are fine because inference gets cheaper, pricing improves, or the product becomes more efficient with scale. Maybe. But those are assumptions about where the business is going, not about its economics today. I want to understand what has to go right to get from one to the other.

The harder question is what the current margin says about the product. High compute costs can come down. Heavy FDE costs are more concerning if customers need those people for the product to work. And for neoclouds, the gross margin line misses too much of the capital structure to tell you much on its own.

What this means if you’re the one presenting the number

Gross margin is no longer a number you should expect an investor to read and move past. Expect them to reconstruct it. If you’ve already done that work, the conversation can move quickly from accounting treatment to the economics of the business.

Before the meeting, I’d have answers to three questions:

1. If someone stripped every dollar of AI cost out of OpEx and put it back into COGS, what would your gross margin be?

2. How much of your expected margin expansion comes from lower unit costs, and how much requires more pricing power?

3. And if your FDEs disappeared tomorrow, would your revenue line move, or just your COGS line?

There’s been chatter that founders and VCs are using inflated ARR to “kingmake” / pick winners on numbers that don’t hold up. Instead of taking the hot take at face value, I asked the people writing checks: what’s the metric you no longer trust, and what’s it usually hiding?

The most popular answer: ARR itself

More people named this than anything else. The term has quietly stopped meaning one thing.

Niko Bonatsos, Founder and Managing Director at Verdict Capital: “ARR... Is it revenue or GMV? Is it recurring? Is it quarterly * 4? Or is it 365 * daily revenue (not kidding - have seen that too!!)?”

Jeff Weinstein, Partner at FJ Labs: “I suspect everyone will coalesce around not believing ARR, which is almost never recurring these days. Every company is now 0-10 or 0-50 in a year. While it’s true there are many unprecedented-growth stories, you need to be skeptical of the numbers and what’s actually being reported.”

Vivek Ramaswami, Partner at Madrona, put a different spin on the same observation: “There are just so many different variations of ARR at this point, so it’s less ‘do I believe your metric’ vs. ‘how do you define what this metric is.’ I encourage founders to include a definition of ARR in their deck, or at least voice over it, so investors know how it’s calculated. Is it actual recurring revenue with durability, or an annualized figure? Both can still be signals of great companies, but it’s helpful to know how to think about the nature of the business.”

Vivek’s suggestion is a good one: A founder who explains their ARR math on slide one builds more credibility than any growth rate can.

The runner-up: CARR

If ARR is bent, CARR (contracted/committed ARR) worries people more.

A Partner at a large crossover fund: “I don’t know why so many companies try to now raise off of a ‘CARR multiple’ instead of an ARR multiple. It’s not a real number... We’ve seen that it doesn’t all convert, and sometimes the drop-off can be quite large, so giving the ‘valuation credit’ seems like a fake way to do valuation. The gap isn’t always predictable either - sometimes customers change their mind, sometimes they just don’t onboard, and sometimes the implementation period can be unpredictable. We would prefer to just hear ARR in the pitch rather than CARR.”

Kevin Tu, Partner at DFJ Growth: “I’ve stopped taking contracted run-rate at face value, particularly for the startups supplying training data and RL environments to frontier AI labs, as it conflates signed bookings with realizable revenue capacity. Eight-figure SOWs signal genuine demand and a supply-side crunch, but are often a poor proxy for revenue: recognition hinges on milestone delivery, customer testing and acceptance, and the vendor’s ability to scale increasingly complex, bespoke work. Because the definition of high-value training data evolves alongside model capabilities, a large backlog may not convert into revenue on a predictable timeline. Ultimately, underwriting the revenue requires going under the hood on both the technology and the scalability of the delivery model.”

Ankit Sud, Partner at NewView Capital, sees the same pattern from a different angle: “Contracted or signed ARR has lost its meaning among AI businesses. This isn’t dissimilar from fintech companies quoting signed payment volume. Instead, we look at live ARR and validate estimated usage ramp by speaking directly to customers about their expectations.”

The consumption-pricing wrinkle

A few responses zeroed in on a specific aspect of how AI products are actually priced.

Francisco Gimenez, Partner at 8VC: “For PLG/consumer, annualized revenue is too misleading. Specifically because so many of these companies are doing blended recurring and consumption pricing models. So you can annualize the recurring monthly revenue – which is still technically wrong unless they are yearly contracts – but you shouldn’t annualize the consumption side until it is truly predictable, which is not usually the case at the early stage.”

Shaun Lee, Partner at Mubadala Capital: “I’ve become cautious of run-rate revenue or ARR for capital-intensive AI infrastructure businesses when it is presented without cash conversion. Revenue may be underpinned by leverage, revenue-sharing arrangements, equipment financing, or large upfront infrastructure commitments, so the headline growth may obscure how much economic value is actually accruing to the company. I want to understand the FCF profile after financing costs and the capital required to deliver that revenue.”

David Roos, Partner at Core Innovation Capital: “Been seeing a lot of ARR annualized off one hot month. Particularly for AI-native services, there’s a lumpiness and uncertainty that needs to be accounted for.“

Gross margin

Everyone expects the ARR pile-on. A few people flagged gross margin:

Sunil Chhaya, Co-Founder & General Partner at Kearny Jackson: “Gross margin, at least for anything compute-heavy that isn’t priced on usage. If pricing is flat but costs scale with usage, that margin number is really just a bet that customers won’t use the product much. We’ve seen founders underestimate compute costs in their GM assumptions over and over, so the number holds up right until the product starts working. Then usage grows, pricing stays put, and margins get squeezed at the exact moment the business succeeds.“

Vivek Krishnamurthy, Partner at Commerce Ventures, flagged a related but distinct version of the same problem: “The metric I actually see equally abused is gross margin. A lot of the AI-native companies I see have an FDE motion, but they push the bulk (or all!) of that cost below the line in OpEx. It only takes a customer call or two to realize that FDE/integration costs scale somewhat linearly with revenue growth for some of these companies. In a world where that FDE cost was accurately portioned, I suspect we’d see a lot fewer ‘>70% GM’ companies!”

The gross margin looks clean because a real cost is hiding somewhere else on the P&L. That’s an uncomfortable thing for a founder to sit with, but it’s a pricing and accounting conversation waiting to happen either way.

NRR

Nikhil Basu Trivedi, Co-Founder & General Partner at Footwork “I’m actually increasingly skeptical of NRR bc we’re seeing a lot of companies start with small POCs and then convert those to contracted ARR, and then include that to produce an amazing NRR metric that doesn’t really feel apples-to-apples with traditional definitions of NRR.”

And the more structural take

A partner at a growth stage fund: “At the growth stage, I don’t think there is as much obfuscation of metrics or worry about companies misleading you. I think the dynamic we are seeing is just that it is hard to do metrics-based investing. Pre-AI, one could look at growth, NRR, GRR, burn - easily benchmark it across companies and then use those metrics as heuristics to invest. In the AI era, you see a lot of incredible growth rates, and it is more about making a qualitative judgment call around durability.“

Founders aren’t squinting more than they used to. The old playbook, benchmark growth, NRR, GRR, burn against last year’s comps, was built for slower growth curves and simpler pricing. AI broke both assumptions, a new playbook has not yet been established.

One more, from a different corner of the market

Everything above is a SaaS story in some sense. But the same failure mode shows up outside software too - it’s just wearing a different metric’s clothes.

Jason Kalira, Partner & Head of Seed Fund at Westly Group:

“Two metrics that jump out to me in physical AI decks specifically, both versions of the same problem – a demo environment standing in for the real world.

The first is autonomy rate. Founders love to show ‘95% task completion, fully autonomous,’ and it’s almost always measured in a controlled pilot with clean lighting, consistent materials, and an engineer standing nearby to intervene. In the time I’ve spent on factory floors with our partners, that number moves a lot once you introduce real-world variability, and nobody’s updating the slide to reflect it. I’d ideally like to know the autonomy rate after 90 days in production, not week one of the pilot. The second is pipeline built from signed LOIs and MOUs, which show up constantly in industrials and energy decks as a stand-in for revenue visibility. Those documents are usually non-binding, carry no penalty for walking away, and often get signed early in a vendor evaluation process, well before budget is allocated or a purchasing decision is final. I want to know how many have a PO or a deposit behind them.

Both come down to the same thing: a number generated under ideal conditions getting presented as if it’s steady-state.”

Swap “autonomy rate” for ARR and “signed LOI” for CARR, and it’s the exact same conversation the SaaS investors above are having. The pattern isn’t specific to any one sector or metric - it’s what happens whenever a number gets measured under ideal conditions and then quoted as if those conditions are permanent.

My takeaway

If you’re a founder reading this: nobody here is hunting for a reason to say no. It comes down to three questions. Is it working? How well? What can you actually extrapolate from that? ARR flavors, CARR, autonomy rate under studio lighting, all decoration, and hype like this gets exposed eventually. Founders don’t get to pick which metric leads the conversation. They get to pick whether they’re managing expectations or building something that survives when the lighting changes.

The shiny new robot that someone spent years creating worked perfectly in the lab, but it toppled over the second it hit the job site. Not a failure of the model – a collision with reality. Concrete expands, Wi-Fi drops behind rebar, and construction noise throws off calibration.

While robotics is not easy, the reality is that the field is much further along than people think. Robots are on construction sites, in factories, and in warehouses right now, doing paid work for real customers. FieldAI just crossed $100 million in revenue and customer contracts - real money from robots working mines and job sites, and raised $405 million to keep deploying. While that “falling robot” image persists in the public imagination, the industry has already moved past it.

That gap between the meme and reality is exactly where the opportunity lies. And since everyone and their mothers are talking about Physical AI right now, here’s what we keep coming back to: the standard software playbook does not quite fit. These are the things that matter.

Bring a Capital Story

$6.4 billion went into Physical AI startups in Q1 2026 alone, across 27 deals. And the capital is flowing for good reason: the technology has genuinely improved. VLAs work, robot arms are cheaper, LiDARs cost a tenth of what they did five years ago.

Hardware timelines differ fundamentally from software cycles. Deploying a robot requires weeks of on-site integration rather than the rapid provisioning of a virtual machine. This is simply the nature of the work.

While a great product story often suffices in software, Physical AI demands a clear capital story: what you’ll raise, when, and what each deployment milestone unlocks. The best founder we’ve heard of walked us through exactly that. That confidence told us more than the demo did.

The Moat Is Deployment

Getting a robot live and learning from it beats any simulation data you can generate. The teams that are great at commercializing and deploying are learning faster than everyone else, every single day. Being on the factory floor, understanding the workflows firsthand, earning the trust of plant managers site by site: that’s a compounding advantage, and it’s available right now to anyone willing to show up.

Whether it’s the Synphony team automating strawberry harvesting, or Generalist and Eka Robotics getting models into real-world hands, the winners are the teams that prioritize shipping over research excellence. Every deployment generates data that the next competitor can’t buy.

Everyone debates the nuances of data collection – egocentric, teleops, and tactile – and the choice between custom- and off-the-shelf hardware. Because so much of data collection is labor-intensive, the specific tech choice matters less than the approach. We’re looking for a clear data strategy that proves your advantage will compound over time.

Work with the Gatekeepers

Siemens, Honeywell, and ABB are established incumbents managing enterprise robotic fleets. As full-stack startups and ‘robot-as-a-service’ models gain ground, the role of these traditional gatekeepers is in flux. The likely winning strategy will include these incumbents in some shape or form - it’s never too early to start building those relationships.

Pilots and the War Chest

Here’s a place where the software rules half-apply. War chests matter in this category. Hardware cycles are long, and if a robotics winter comes, capital is what carries you through it. Pilots are great - but we’re so early in the cycle that it’s hard to read into the revenue unless there are customers ready to order 1000s of robots.

But indigestion is real, too. Too much money too early leads to overstaffing and over-equipping before you understand the environment you’re deploying into. The tell is headcount growing ahead of deployed revenue; it shows up before the burn does. The healthiest pattern we see: raise for resilience, but let milestones set the pace. Unit economics do ultimately matter, as we can’t transform the world without it - and a large war chest can give you time to subsidize things, but that can’t be the crutch.

While startups can raise too much and cause indigestion, physical AI milestones are more than just hitting 3x ARR with nice SaaS metrics.

What Actually Scales

This category has no shortage of brilliant engineers. What it’s short on: founders who can walk into a facility, win over a plant manager who has sat through ten robotics demos, and structure a deal that survives procurement.

Add the grit to keep operating through real-world setbacks and the pull to recruit engineers who’ll happily show up to a job site at 6 am, and you have the rarest profile in the category. Worth leaning into hard when you find it. In diligence, that means spending as much time with the commercial founder as the technical one, and asking who closed the last three deals.

Where the Winners Are Emerging

Winners are emerging where labor is expensive, and the customer has real buying power. Industrial use cases from warehouse to manufacturing to chemicals and logistics - structured environments drive rapid data accumulation and unit volume at scale.

The less obvious opportunity lies in the middle of the stack: simulation infrastructure, sensing, and operational tooling. Cloud ran the same play: the headlines went to the apps, but Datadog and Snowflake won in the middle. Physical AI will produce the same kind of quiet platform winners.

And on the headline products themselves: vertically integrated robotic companies that customers can train on new tasks is the decade-long bet; narrow vertical applications are the three-year story. Both can be true, but they have different risk profiles and may not be the same company. Whether a given capability ends up as a feature or a platform depends on how fast vertical integration moves — and it’s genuinely too early to call.

The Long Game

The hardware is finally ready, and the models are capable enough to matter. Real-world deployments are generating actual revenue today, proving that the old “falling robot” caricature is well out of date.

The industry’s primary metric will shift from ‘total operational hours’ to ‘number of robots per customer’.

Scaling into these environments — where everything is tested, and nothing is guaranteed — is the real challenge. It takes years, not quarters, but that’s exactly why the companies deploying now are building something truly durable. The teams showing up to do the hard work today are earning their place in the room. The failing robot meme belongs in a different era.

This past April, HubSpot started charging fifty cents for every conversation their AI actually resolves. They aren’t the first, either. A few others, including Zendesk and Intercom have already headed in that same direction.

When you stop and think about it, that’s a huge shift. You aren’t just paying to use a tool anymore; you’re paying because a customer’s problem actually got fixed. Software is moving away from just selling you access and moving toward taking real responsibility for the results.

According to the Gartner/Deloitte TMT Predictions 2026, per-seat SaaS models dropped from 21% to 15% of the market over the last year. That’s not just some rounding error – the market is actively repricing what software is actually worth.

And it’s signaling a much bigger shift.

The Chain Reaction

AI goes way beyond just cutting dev costs; it’s a fundamental reset. AI amplification of human output does more than increase speed; it unlocks the ability to personalize at scale, leading to outcome ownership, better value capture, and ultimately, the rise of vertically integrated businesses.

As software takes on more of the heavy lifting, the economics of the value chain are fundamentally shifting. Companies can move closer to the customer’s end goal, capture more of the value they create, and in some instances, reinvent industries from the ground up. The recent change in HubSpot’s pricing is just the starting gun for this shift.

Regulation Is the Moat

Regulated industries are where the best AI bets are hiding. Sales cycles are brutal. That friction compounds into an advantage — once you’ve earned clinical or fiduciary trust, retention is structural, not contractual.

Look at Range in wealth management. Clients aren’t buying software; they’re handing over their financial lives. Sure, the sales cycles are brutal, but that friction is actually a competitive advantage. Once you’ve earned that kind of trust, you’re not going anywhere. You’re sticky in a way standard SaaS never will be.

It’s the same with Qualified Health. Their advantage isn’t a ‘better’ model—everyone has the same tech now. They win because hospital systems trust them with real clinical work where the stakes are life and death, not just data entry.

Stop seeing compliance as a burden. It’s your barrier to entry. AI might write code faster, but it can’t accelerate trust. That’s the real work, and that’s how you build a business that actually lasts.

Customers Want Results, Not Software

CX platforms like Sierra and Decagon sell outcomes. The software is how they deliver; the result is what they charge for. The question customers are asking has changed. Not ‘what does the software do’ but ‘did my problem get solved, and what did it cost me. The software is just the plumbing; the result is what they’re paying for.

HubSpot’s pricing shift is the writing on the wall. When you charge per resolution, you’re ditching the ‘tool’ label and taking responsibility for the outcome.

Two businesses can have the exact same revenue, but entirely different futures depending on how close they are to the money. One just gives you a hammer; the other builds the house. That difference is everything.

The Real Metric: Value Capture

Standard metrics like ARR and NDR are fine for keeping score, but they aren’t the whole story for AI-native companies.

If you save a client $1B and take 1% of it, you’ve got a very different business than someone who saves $1B and captures 20%. The question isn’t ‘what’s your ARR?’ It’s ‘what percentage of the value you created are you actually keeping?’

Take legal services. The standard move is selling a tool to law firms. But the AI-native move asks: if technology is doing the heavy lifting, why not just be the law firm?

Firms like Crosby are asking whether we should rethink the institution itself, not just the software it runs on. We should stop asking ‘how do we help lawyers work faster?’ and start asking ‘who owns the client relationship when the software does the work?’

What This Means for the Future

Venture capital used to love ‘asset-light’ businesses because adding service or customization meant hiring more people and killing your margins. AI changes the math.

Because productivity is skyrocketing, companies can go much deeper into the workflow and deliver the outcome themselves without ballooning their headcount. The old arguments about scaling and integration no longer hold water.

The winners of the next decade won’t look like the SaaS giants of the last one. They’ll be right in the thick of the workflow, vertically integrated, and capturing much more value than anyone thought possible.

Who in your industry is closest to owning the outcome right now - and what’s stopping everyone else from getting there?

Last week, all eyes were on SpaceX’s IPO, which priced $135, opened at $150, and didn’t look back. And this is just the beginning! Anthropic filed confidentially two weeks ago, and many expect OpenAI to IPO toward the end of the year.

So what does this mean?

The public markets’ new benchmark

The window isn’t just open, SpaceX’s offering was absorbed cleanly, proving the capital is there. For those who spent the last two years wondering if the public markets could handle this generation of companies, we have our answer.

Is it time to ring the bell?

Databricks just closed another private round, Series M(?), if you’re counting - and Stripe is in no hurry. When private capital still flows freely and your comps just cleared $2 trillion, why deal with the quarterly earnings circus? My read: several more big names stay private longer than expected, sustained by LP capital recycled from the very exits everyone is celebrating this week.

Beyond the software stack

Physical AI (robots), supply chain resilience, quantum, there was even a quantum computing (Quantinuum) IPO recently that priced well. It’s a major structural shift to see these sectors finally attracting serious money from both private and public markets. That said, this is also the point in the cycle where caution is warranted: all it takes is one high-profile failure, and the sentiment reverses fast. The window is open, but not permanently. These aren’t companies with similar ARR profiles to the labs.

The DPI wave

It will replenish the coffers and fund the next SpaceX. But lockup periods mean the cash reentry is considerably slower than the headlines suggest. LPs won’t be immediately flush. The planning cycle starts now, but the actual capital deployment is a 2027-2028 story. Founders thinking they’ll close a seed round off SpaceX buzz in Q3 should temper expectations.

The new calculus for fund performance

Firms with meaningful early AI exposure, whether through direct positions, secondaries, or portfolio companies that rode the wave, will raise their next funds on exceptional DPI numbers.

That performance unlocks larger funds, which buy access to the best deals at better terms, which produce the next set of strong numbers. The long tail of newer managers will face real LP pressure unless they were genuinely integral to the AI story.

The widening divide

Returns are concentrating on the platforms and insiders who had prior access. Those firms walk away from this cycle with structural advantages that are very hard to close. Everyone else is largely a bystander to the upside on these specific names, even if they hold public shares. The barbell widens - mega funds and the most differentiated emerging managers on one end, a very compressed middle on the other.

Let’s be clear: the SpaceX IPO doesn’t mean the floodgates are open for everyone; it’s a win for the heavyweights who were already well-positioned. The real test is where that recycled capital goes next – whether this capital cycle produces the next generation of ideas, or just pays higher premiums for more of the same?

Plot twist: this week’s post is different. We typically stick to signals and market analysis, but after hosting an Agentic Finance Summit with 150 people actively deploying AI in financial services, we had to share what we actually learned.

What emerged was a clear pattern that contradicts much of the market noise: the companies winning in enterprise AI are the ones solving the infrastructure data problem. And most vendors don’t even know what that problem is. What follows are some key takeaways from the conversations.

What Vendors Often Get Wrong

“A lot of these vendors and especially young startups tend to focus a lot on just kind of going too broad in terms of the sort of the problems that they want to go solve for and they don’t appear to be very focused on the specific problem they want to solve.” - from a Director of AI at a large financial institution

Most vendors pitch transformation, but enterprises live in complexity. A typical financial institution has decades of acquisitions, migrations, legacy systems, and mainframes running alongside cloud infrastructure. APIs return data that looks clean but contains hidden fraud codes, legacy error messages, and fields that mean something completely different from what their names suggest.

The hard part isn’t deploying AI. It’s making sense of data quality issues that prevent AI from being useful. Vendors building for an ideal enterprise fail immediately. The ones getting traction come with a prescriptive thesis: “Here are three specific problems we solve for you in KYC or AML within wealth management.”

The Real Moat: Three Layers

Layer 1: Process Intelligence Observe how work actually happens before deploying AI to change it. Create a “digital twin” of core operations. Only then deploy agents reliably.

Layer 2: Compliance-First Architecture Design for governance from day one, not as an afterthought. When you build AI for auditability and explainability from the start, compliance becomes a foundation, not friction. You get systems where every interaction is monitored, where policy changes can be tested offline, and where you can A/B test approaches.

Layer 3: Systems Modernization Use AI to accelerate the modernization of legacy systems. Compress what normally takes years into weeks.

The Beachhead Playbook

Success requires going deep on one wedge before going wide. Start with a specific problem, build a solution that scales, and respect governance.

What Makes Pilots Succeed:

• Use real production data, not sandboxes

• Get to UAT environment, production-ready

• Pick one use case and demonstrate mastery there

The Three Boxes to Check: Prove the tech works with their data in their environment. Do it resource-efficiently for both parties. Ensure genuine executive sponsorship—someone whose career is better if the pilot succeeds, not someone checking a box for board pressure.

The Wrong Play: Start Simple. Conventional wisdom says start with low-hanging fruit. This is backward. Proving you can pick easy problems doesn’t prove you can handle hard ones. Start with a genuinely complex process. Build infrastructure that handles complexity. Easy wins come later.

When pilots actually move into broader deployment, they’re the ones who went after legitimately difficult cases. They proved something real.

What Everyone Gets Wrong

Misconception #1: One-Shot Deployment AI deployment is iterative. Month one might be 30%, month two 60%, approaching 100% over time. If you judge on month-one performance or aim for 100% from day one, you’ll fail.

Misconception #2: Accuracy as Success The fixation on 99% accuracy stops deployments prematurely. The question should be: what did you learn? What applies elsewhere? 90% accuracy with human-in-the-loop is often good enough to go live.

Misconception #3: Agents Are Too Risky for Regulated Work Compliance teams see hallucination and drift risk. These are real. They’re not blockers, they’re design problems. With proper guardrails and monitoring, you can manage the risk.

What This Means

For founders: pick one part of the enterprise AI problem and solve it completely. Understand the messy reality your customer lives in. Build for that reality, not the ideal. Get one wedge right with real data and real stakes.

For enterprises: you’re not looking for the best AI. You’re looking for the deepest understanding of your infrastructure problem. Does this vendor understand your legacy systems? Your governance? Your actual workflows? Have they solved this for companies like yours, or are they applying a generic solution?

The vendors who deploy are the ones who understand infrastructure. Everyone else stays in the pilot sandbox.

The Budget Crisis Is a Feature, Not a Bug

SaaS trained companies to treat usage as a proxy for value. AI breaks that metric, as the same workflow can consume radically different amounts of inference depending on factors like the prompt, context, or model used.

At an event we held last week, one founder put it directly: “We looked at our one-year roadmap. We asked ourselves: Do we think we could get this done in six months? Is that worth $100 million in tokens?”

The answer was obvious. Yes.

70%+ of committed code now touches an AI model at companies using these tools aggressively. Engineers who were skeptical months ago are now shipping twice as fast. The budget broke because the tool worked so well that success arrived faster than anyone forecasted.

This is the strongest bull signal, not waste.

Explosive Adoption Is Proof the Tool Works

When 84–95% of your engineering team is actively using a new tool, that’s not a budget problem -- it’s the clearest product-market-fit signal you can get.

Engineers used Claude aggressively because it saved them time and delivered more than predicted, proving its value (product-market-fit) and marking the true start of the business, rather than simple token maxxing.

Today’s tools are the floor, not the ceiling, which means the ROI you’re seeing now is the minimum case, not the benchmark. Optimizing around current limitations misses the compounding effect of rapidly improving capability.

We Are Watching Healthy Jevons Paradox Play Out

This happened before. During my time at VMware, as the cloud was arriving, enterprises initially spent more on infrastructure, not less. Cheap compute made new work economically viable. Backlogs cleared. Volume expanded.

Anthropic’s run-rate jumped from $9B to $30B. The demand curve revealing itself. We’re at roughly 1% of AI penetration in enterprise, watching the beginning of adoption, not the end.

Companies that won in cloud kept investing through the “wasteful” early phase. The ones that throttled spend lost ground.

The mistake is assuming that every dollar of AI spend should immediately appear in business outcomes. Most organizations have only accelerated one layer of the stack. Engineering may move twice as fast, while approvals, compliance, go-to-market, procurement, and decision-making still operate at yesterday’s speed.

Early spend often shows up first as velocity. The outcomes arrive later.

Don’t Stifle Imagination in the Gold-Rush Phase

Governance will come. It should. But if you’re optimizing for efficiency before you’ve discovered what’s possible, you’ve already made the more expensive mistake.

AI isn’t coming for jobs first. It’s coming for the instinct to optimize before discovering what’s possible.

The best breakthroughs are still being discovered. Rigid spend limits kill optionality right when optionality is most valuable. Most companies won’t find what’s possible until they let their best people play without guardrails.

Once you know what works, optimize. But optimize before discovery and you’ve already lost.

Treat early overspend like R&D acceleration, because our imagination is too small for what is to come.

Tokenmaxxing Done Right Becomes Outcome Maxxing

Tokenmaxxing without measurement is waste. With obsessive outcome tracking, it’s acceleration.

Volume leaderboards and unchecked consumption are governance issues, not proof the tech is broken. Microsoft consolidating Claude licenses wasn’t rejection, it was a platform play after value proved itself.

Winners will layer smart constraints on raw spend:

• Spend on what actually compounds

• Measure what experimentation produces

• Track who moved the needle

• Optimize for velocity before optimizing for efficiency

The goal is not lower token spend, the goal is higher output per dollar spent - those are not the same thing. Winners will layer smart constraints on raw spend:

• Spend on what actually compounds

• Measure what experimentation produces

• Track who moved the needle

But here’s what changes the equation: open-source models shipping out of Asia are already running at 80-90% quality of frontier models. The closed labs are subsidizing usage, so price doesn’t force the comparison yet. When labs have to charge a real margin, open source becomes a serious value alternative. The playing field levels.

The goal is higher output per dollar spent, not lower token spend. Those are not the same thing. And as alternatives mature, that calculus shifts.

Who Owns the Acceleration?

Those who treat it as a spending problem miss the insight: they’ve found a tool that works so well their teams can’t stop using it.

Companies that throttled cloud spend in 2012 lost ground they never recovered. This moment is the same shape.

Who owns moving the needle with AI in your organization - and how are you measuring whether your imagination is actually being realized?

The Bill Just Arrived.

Here’s what just happened: Corporations are realizing they’ve blown through their entire AI budgets for 2026. Not next year. This year.

Coincidentally, OpenAI just gave away two months of free Codex access.

Think about that move.

The invisible cost era is ending.

AI spend used to hide in small line items - SaaS seats, T&E, experimentation budgets. Now it’s becoming a real P&L line. CFOs are starting to ask: What are we actually getting for this?

Most companies don’t have a good answer. And now they are running out of budget while trying to figure it out.

The Real Problem is Measurement.

You can see who has access to AI. You can’t see what they’re producing with it. The person using 3k tokens and the person using 10k tokens look identical in a budget report.

One might be your most productive employee. One might be your most wasteful. Without understanding output, you don’t know which is which.

Why Cloud Optimization Won’t Work Here.

Cloud optimization was effective because the unit of waste was clear: idle compute. You could see it, which meant you could cut it.

AI is fundamentally different. The unit of value is ambiguous, so cutting the wrong usage might destroy productivity you didn’t know you had.

That’s where Jevons paradox makes things harder. As AI gets cheaper, more work becomes worth doing. The backlog clears. New work appears. Volume expands. That’s a good thing.

But it creates a new problem:

Cost is the Wrong Frame.

So this leaves you wondering: What does good output actually look like?

This is where the org breaks. Most companies are pushing AI adoption without changing how they operate. If AI can do what three people used to do, you have three options:

• reduce headcount

• redeploy people

• or let them stay busy doing work AI is already doing

Most companies default to the third.

Being AI-Native Means Restructuring, Not Just Adding Tools.

Becoming ‘AI-native’ means restructuring how work gets done. Most companies are skipping that part. This creates a new management problem. You’re no longer managing just people. You’re managing a system of humans and agents. Agents don’t have performance reviews. They don’t push back. They don’t admit when the task was wrong.

That means oversight becomes the critical function:

• what is being done

• whether it’s working

• when to intervene

The Test is Happening Right Now.

OpenAI’s free offer is a forced reveal: would enterprises pay for a better model, or take the cheaper one?

The answer many enterprises are about to give will tell the market a lot about the future of AI pricing and feature value. If enterprises collectively say “we’ll take the cheaper option,” that’s a signal. If they say “we need the better model, cost be damned,“ that’s a different signal.

Who Owns This?

The AI bill is coming either way.

The companies that treat it as a cost problem will optimize for cost and destroy hidden productivity. Those who treat it as a management problem will establish clear accountability and build an operational structure that continually improves.

On May 4, 2026, two announcements landed within hours of each other. Anthropic closed a $1.5B joint venture with Blackstone, Hellman & Friedman, Goldman Sachs, Apollo, and General Atlantic. Hours later, OpenAI unveiled a roughly $10B Deployment Company pulling in TPG, Brookfield, Bain, Advent, SoftBank, Dragoneer, and nineteen other firms. The speed and the symmetry weren’t a coincidence.

Most coverage framed this as an AI distribution win. It’s actually two trades. The labs got distribution. PE got a credible operational story to bring to LPs at exactly the wrong moment to be without one. DPI, the cash actually returned to investors, has collapsed to 6% of PE AUM in the twelve months ending June 2025, against a 2015-2019 average of 16%, according to McKinsey’s 2026 Global Private Markets Report. Financial engineering isn’t generating returns the way it did. The labs aren’t just selling intelligence. They’re selling sponsors a way to prove the operating side still works.

The scale is the part most people miss. Since 2000, US PE-backed companies have grown from roughly 2,000 to over 11,500, a 400%+ increase, while listed companies on NYSE and NASDAQ shrank from roughly 7,000 to 4,500. Winning one PE fund isn’t a single enterprise sale. It’s a distribution mandate across a whole portfolio, enforced through operating partners and 100-day plans that now carry explicit automation targets. The labs don’t pitch each portfolio company. They embed at the fund level, and adoption follows.

The demand is real, but the gap is wide. A 2026 FTI Consulting survey of 200 PE fund and operating leaders found that while most report positive early impact from AI, just 7% of portfolio companies have reached enterprise-scale deployment. That gap is the JV’s reason to exist.

Why mid-market PE is the customer the labs actually want

Fortune 500s negotiate. With data centers, security teams, and real leverage, they push for on-prem or custom hosting -- and every concession they win compresses the labs’ recurring revenue.

PE-backed mid-market companies can’t play that game.

Regional banks, mid-sized manufacturers, healthcare groups, restoration businesses: no GPU clusters, no ML ops teams, no capex appetite for self-hosting. On-prem only pencils out at token volumes they’re unlikely to hit for years, if ever. So they default to cloud APIs—the only viable path, and exactly where the labs want them: usage-based billing, full pricing power, no integrator in between.

A PE portfolio company calling the API has a relationship with the lab, not with CoreWeave or AWS. That’s where the stickiness, pricing power, and recurring revenue live. Hyperscalers pushing harder into the model layer could shift that, but today the customer belongs to the lab, and thousands of sticky mid-market contracts are the kind of revenue an S-1 gets built around.

The timing isn’t accidental.

Goldman Sachs’ “Decoding the Agentic Economy” (May 5, 2026) lays out a forecast worth taking seriously, not for the precise numbers, but the directional logic. Token prices used to fall about 40% a year. That trend has stalled, even reversed, while compute costs keep dropping. The spread between what the labs charge and what it costs them to serve a token is widening for the first time. Goldman argues margins are improving right now, in the first half of 2026. The labs are locking in volume at the exact moment volume gets profitable.

Goldman models a 24x rise in token consumption by 2030 – not because more people are using AI, but because they’re using it differently. A chatbot session burns roughly 1,000 tokens; a persistent agent running across business workflows could burn 100,000 per user per day. That’s a 100x jump in intensity per seat before broader adoption is factored in. PE portfolio companies onboarded today are the first customers for a much larger wave.

What founders should take from this

For founders selling into the same mid-market, the JVs come with structural advantages that are hard to race against:

• Direct engineering deployment from the model provider – the implementation relationship is captured before you can pitch it.

• A top-down portfolio mandate that doesn’t require a sales cycle – the fund’s operating partners enforce adoption.

The positions that hold up are specific. Hyper-vertical software with genuine domain advantages, compliance requirements, or trust barriers that generalist teams can’t match. Pure software plays work if your value sits above the implementation layer – where the JV deploying the model actually sends customers your way rather than taking your spot. And infrastructure the JVs need to function: the process intelligence and governance layers that have to exist before any of this works in production.

The JVs don’t need to win a sales cycle against you – they’re already inside the portfolio. Reposition now, before the JVs have two years of embedded relationships and the mandate is locked in.

The opportunity for founders is everything the JVs can’t parachute in and replicate in 8 weeks: domain expertise, the compliance context, data that exists only in a specific vertical.

Cognition, Cursor, Decagon, Hippocratic, EvenUp, and Intercom - six vertical AI companies. All announced their own models within days of each other. The narrative was identical: we have proprietary data, we trained our own model, that’s our moat.

The debate this kicked off is familiar. Vertical models vs. general models. Proprietary data vs. frontier capability. Who wins?

The vertical model argument isn’t irrational. Smaller models are faster and cheaper to run. In high-volume support or coding environments, latency and inference cost aren’t abstractions; they show up in unit economics. And in regulated industries, the choice isn’t always strategic. A hospital sending patient data to a frontier API may not have a legal option. A bank routing client conversations through a third-party model may be violating its own compliance policies before it even gets to the moat question. For these companies, building a vertical model isn’t a bet on data superiority. It’s a requirement. But compliance solves a constraint. It doesn’t build a moat.

So the vertical instinct isn’t wrong. The problem is what it optimizes for.

Fine-tuned model advantages get wiped out every few months when the next frontier release ships. Cursor is the clearest recent example. They built a meaningful coding advantage on top of Kimi K2.5, two years of professional engineers correcting real code inside private systems. Anthropic and OpenAI closed that gap on complex coding benchmarks without ever touching a private codebase. The “when” still matters -- a company might get 18 months of differentiation or 90 days. Nobody knows in advance. But the pattern is consistent enough that model choice should be treated as a depreciating asset rather than a durable one. The companies racing hardest on it right now are optimizing for something that resets.

AI is not tooling. It’s labor. It’s infrastructure. Tooling doesn’t get better at your company specifically. A new hire does. The difference in frame changes every decision that follows -- how you evaluate it, how you expand it, what you measure. It’s also what’s driving a growing problem in board meetings. AI spend per employee keeps climbing. Productivity gains aren’t showing up at the same rate. Even Meta, which cut 10% of its workforce to shift spend toward AI, is an outlier. Most companies are paying for both -- the humans and the AI -- without seeing the productivity delta that justifies it.

Soon enough, CFOs will step in. Not to champion AI, but to demand proof that it’s working. The companies that can’t answer that question with real numbers will see budgets cut before the system ever gets a chance to compound. Companies treating AI like software are stacking costs on top of existing headcount. Companies treating it like labor are asking a different question: what does this system own, and what can I stop paying a human to do?

What actually matters is what happens to the work. Customer support is a good example. The L1 / L2 / L3 structure wasn’t designed; it was a workaround for scarce human attention. AI breaks that constraint.

• L1 is going

• L2 is next

• L3 is closer than most support leaders want to admit

The collapse is happening more slowly than the technology allows. Vendor incentives point the wrong way -- consumption-based pricing rewards more tokens spent, not fewer tickets handled. And most buyers haven’t written RFPs that hold vendors accountable to deflection rates or resolution metrics. When CFOs start demanding ROI, both dynamics change fast.

Whether a vertical model or a frontier model gets you to 90% automation first is almost irrelevant. The structure collapses either way. The org chart doesn’t flip overnight. But the direction is clear. The pyramid of human agents with AI tools on top inverts. What you’re left with is a small oversight team managing an invisible execution layer below them.

So how do you treat the system once it’s in place?

Think of it like an employee. A new hire might have higher raw capability, but your tenured employee is more effective because they understand your customers, edge cases, and history.

We’re already seeing this in the numbers. Intercom’s Fin jumped from 68% to 75% resolution rates for one customer overnight, not from a product overhaul, but from a new model trained on billions of proprietary customer service interactions Fin had accumulated over time. You could read that as Intercom simply shipping a better model. But a better model trained on what? The proprietary interaction data Fin had been accumulating. Data was the real competitive advantage here; the model simply tapped into it to drive better results rather than just lower costs. And it only happens if you’ve been treating the system like something that learns, not just something that runs.

A model gives you the raw skill, but your own data and feedback are what give it real experience. And unlike humans, that context doesn’t leave. What’s harder to transfer is the institutional layer on top—the evals, the edge-case library, the knowledge of what good looks like. That’s what compounds. That’s what’s genuinely hard to replicate.

The process begins by starting narrow, then expanding as trust builds. Trust isn’t a feeling; it’s a quantifiable metric. You expand scope when the system is hitting a defined accuracy threshold on the current tier consistently -- not once, not on a good week, but over enough volume that the edge cases have had a chance to surface. In support, that means starting with FAQs and order status. When accuracy is consistently above your threshold and CSAT isn’t dropping, you move to billing disputes and account exceptions. When that tier stabilizes, you move to escalations end-to-end. The failure mode is moving before the numbers say you’re ready. A billing dispute handled badly costs you a customer. An escalation handled badly costs you a customer and a lawsuit. Scope should follow performance, not enthusiasm.

Change management is just as important as the technical layer. Expanding scope isn’t just a product decision. It’s an organizational one. The teams whose roles are shifting need to be brought along, not surprised.

Once this happens, support stops being a differentiator as costs compress and quality equalizes, turning it into table stakes where competition moves upstream to product, pricing, and individuality. AI can copy taste. It can’t copy an authentic voice. Taste is a style. Voice is a record. You can’t train on something that hasn’t happened yet.

Compliance and cost savings are valid reasons to build your own model, but that model is still the most fleeting part of your strategy. The frontier will catch up. Open-source will commoditize the base. Data portability is improving fast -- the switching cost of moving your accumulated context to a new model or provider is lower than most vendors want to admit. What won’t be replicated is the institutional layer on top of it. The evals. The edge case library. The accumulated judgment of what good looks like in your specific context. That’s the asset. The model just runs on top of it. And the market won’t be winner-takes-all. Cost will be the next selection criterion. As inference prices fall, the right model is increasingly the cheapest one that hits your accuracy threshold. Just like not all employees are created equal, not all solutions have to be. What matters isn’t which model you picked. It’s whether you built the institutional layer on top of it.

The debate between vertical and general models is a distraction from what truly matters for your strategy. Ultimately, both paths can lead to success. The imperative is to regard AI not as a tool but as an integral teammate, and proactively expand its scope until it’s handling every task you are prepared to delegate.

If you’re managing a fleet of humans and agents and thinking through how scope expands — I’d like to hear how you’re doing it!

A month before SpaceX announced a deal to potentially acquire Cursor for $60 billion, Elon tweeted three words about Cursor’s supposedly proprietary AI model: “Yeah, it’s Kimi 2.5.” An open-source base from Moonshot AI, free to anyone. What Cursor had built on top wasn’t: two years of professional engineers correcting real code inside private company systems. Cursor took Kimi K2.5 as a starting point, then put it through continued pretraining and reinforcement learning at 4x the usual scale, a process where the model learns from real-world coding outcomes. Roughly three-quarters of the computing power spent building the final model came from Cursor’s own training on top, not from the Kimi base.

That’s the Cursor product. Not novel software. It’s built on top of Microsoft’s free VS Code editor with AI suggestions layered on top, and the interface could be replicated by any well-funded team. What can’t be replicated are millions of moments where a senior engineer at a real company accepted, rejected, or rewrote an AI suggestion inside a private, high-complexity codebase. That’s not synthetic benchmark data. That’s ground-truth signal from the highest-skill humans on earth, in contexts the labs will never see.

Coding isn’t unique. The same dynamic plays out in any domain where expert humans interact with AI in high-stakes, private contexts: legal matters protected by privilege, clinical encounters protected by HIPAA, financial decisions that never touch a public dataset.

But here’s the harder question: is the last mile actually defensible? Or can the frontier labs just hire their way to the same place?

Two forces decide the answer.

First, signal: does usage produce proprietary data that makes your model better in ways the labs can’t replicate?

Second, switching cost: would replacing you be painful for the customer, regardless of how good the alternative model gets?

Most vertical AI companies have one of these. The companies that survive have both.

No lead is safe

The labs didn't wait for Cursor's data. Anthropic, OpenAI and others hired engineers, built evaluation frameworks, ran synthetic benchmarks at scale, and threw enormous computing power at coding until the models were excellent. The result is that the leading coding models from Anthropic and OpenAI are already genuinely good at the kinds of complex engineering problems Cursor was built to solve, without ever seeing someone's codebase. That gap is real but closing.

The honest version of this is harder than most commentary admits. GitHub Copilot has millions of users and recently began using their interaction data to train its models. Google has enterprise workspace data. Microsoft has Teams and LinkedIn. The labs are not starting from zero in a private context; they have their own flywheels running. The argument isn't that the labs have no signal. It's that structurally protected data - inside a bank's risk systems, inside a hospital's EHR, inside a law firm's active cases - is the one category the labs genuinely cannot access, regardless of how long they run their own flywheels.

The labs can approximate the signal. They can’t perfectly replicate it, but good enough is all they need to make the switching argument collapse for most users.

The vertical AI moat isn’t “we have expert signal and the labs don’t.” It’s narrower: the only durable moat is structurally private context, data that is legally or practically impossible to replicate (not just hard to get, but structurally off-limits), regardless of how many engineers a lab hires. Cursor's traces from inside proprietary systems at banks and financial institutions. Range‘s data covering a client's full financial life in ways no public dataset captures. Qualified Health’s clinical encounter data, captured through an AI layer embedded directly in hospital workflows and protected by HIPAA. Some of it the labs will eventually approximate. The data under legal privilege or regulatory protection they never will.

Even that head start has a clock on it. A better model is the first thing you build with it, but model advantages erode. The thing that matters is becoming so embedded in how a team works, how a firm operates, how a codebase grows, that replacing you has nothing to do with whether a better model exists.

Open source doesn’t threaten this. It completes it

Most commentary treats open source as a wildcard that might disrupt the vertical AI story. That’s the wrong frame. Open source is the best thing that could happen to vertical AI companies, if they’re positioned correctly.

The closed-source labs’ real advantage was never the base model architecture. It was compute and post-training data. Once the base model layer becomes freely available (and Meta’s open-source Llama models are already competitive on most everyday tasks) the labs lose the lever they’ve been using to extract rent from every company building on their APIs. Rate limits, access restrictions, supplier-becoming-competitor dynamics: all of that disappears when the base model is free. The caveat: this assumes Meta keeps releasing competitive open weights, which recent reporting suggests is not guaranteed forever as the company develops proprietary frontier models alongside its open ones.

What’s left?

Two things: compute and proprietary post-training signal, meaning the specialized data built on top of the base model. Compute is expensive but rentable. You don’t need to own Colossus (xAI’s supercomputer); you need access to enough GPUs to run the training. Proprietary signal is the only input that’s genuinely scarce. A vertical AI company with irreplaceable domain data can take an open base model, train on its proprietary corpus, and build a specialist that beats the general model on its domain. No API costs, no access risk, no supplier that’s also a competitor.

The Cursor deal looks different through this lens. It’s not proof that vertical AI companies need a hyperscaler to survive. It’s proof that Cursor had accumulated enough signal to make a large-scale training run worth doing, and needed the compute to execute it. If open base models get good enough and GPU rental costs keep falling, the next Cursor might not need SpaceX at all.

The caveat is compute cost. A meaningful post-training run at the scale that actually moves the needle is still prohibitively expensive for most vertical AI startups, which is precisely why Cursor needed SpaceX in the first place. Open source lowers the floor on base model access. It doesn’t yet lower the ceiling on what it costs to do something significant with proprietary signal. That changes as GPU rental costs fall, but the window between “open source is good enough” and “training runs are affordable” is where most vertical AI companies are right now, and it’s an uncomfortable place to be.

The scarce input was never the model. It was always the data to train it on.

The Cursor deal is a bet on time. SpaceX is paying $10B to find out whether two years of professional engineers correcting real code inside private systems is worth more than what the frontier labs can brute-force with compute. Every vertical AI company is running the same experiment right now.

I keep having the same conversation - with investors, founders, corporates, press: what’s next this year?

The macro environment feels tough. IPOs feel stalled. The M&A announcements of 2025 felt like forever in this AI world. The instinct is to assume we’re just waiting for things to reopen. That’s not what’s happening. This is not a paused market. It’s a selective reset.

Capital is still there. But it’s being reallocated much more aggressively.

Investors are concentrating on:

• AI infrastructure

• semiconductors

• clear category leaders

At the same time, much of the software world is still working through a sharp multiple reset. That tension is about to be tested.

If the expected SpaceX IPO does happen this early summer, it could help reset demand and re-anchor risk appetite. But it will also pull capital and attention away from everyone else a few weeks before and after. Even then, if the window opens, it won’t open evenly.

This leads to a harder truth.

This is now a grow-or-fade market.

If you’re not growing, you’re slowly becoming less relevant. AI budgets are expanding across categories. But they’re not being spread evenly.

Profitability alone doesn’t solve this. It can buy time. It does not determine where the spending goes.

The market is rewarding companies that are capturing new budget, not just protecting margins.

The opportunity is that growth is still there. According to Gartner, IT budgets are growing over 10% YoY, but AI spend is over 80%. But vendor count is not. That means growth is coming from share capture, not market expansion.

To win, you have to capture the AI budget in your category. That requires more than layering AI onto an existing workflow. It requires becoming core to how that workflow gets executed. That means LABOR.

In many categories, this will concentrate into a small number of winners. One or two platforms that define the “agent layer,” with everyone else competing for whatever budget the winners don’t capture.

So no, we’re not just waiting.

The companies that can show durable, AI-driven growth will have options: IPO, M&A, or staying private.

The ones that can’t will find that every path to liquidity remains narrow.

AI agents are often framed as replacing traditional software. We’re squarely moving from co-pilot to task agents, but processes will start collapsing as we redefine SOP (standard operating procedure).

Interfaces disappear. Work becomes conversational. Systems orchestrate themselves. That narrative is directionally right.

But it misses where value actually accrues.

Agents do not remove the execution boundary. They expose it.

Reducing UI friction by adding a prompt box does not remove the need for permissioned execution - we’ve all worried about our AI going amok.

The moment an agent takes action, moves money, updates a system of record, or triggers a real-world process, the problem changes.

It is no longer about capability. It is about accountability.

Who’s responsible when it fails? Who controls permissions, approval policies, and risk?

These questions become more important as autonomy increases.

Rippling is a useful example. The product is not valuable because it recommends actions to HR or IT teams. It is valuable because it executes those actions — payroll, device management, identity — within a system that enforces policy, tracks changes, and maintains accountability.

Execution is the product.

The more actions an agent can take, the larger the blast radius when something goes wrong. As we’ve seen with the number of cyberattacks (Mercor, Anthropic, etc.) people’s need to find a simple solution has caused issues that bring the importance of identity systems, policy engines, auditability, monitoring, rollback, and exception handling to the forefront.

These are not secondary features. They are the foundation of execution.

If you want to understand where value sits in an agent-driven world, ask:

• Who controls write access to systems of record?

• Who defines escalation and approval policies?

• Who absorbs risk when something fails?

• Is the system governed or just an overlay?

Action without write-back authority is an assistant.
Action with governed execution is a business.

Governance for multi-player is still in its infancy. Excited to see how that evolves over time as we see the leap from automating tasks to processes collapsing.

Who’s on the hook if the agent fails?

→ Part I: Customer Two Is the Only Metric That Matters

→ Part II: Services Aren’t the Sin. Linear Scaling Is.

→ Part III: Automating Work Is Not a Moat

→ Part IV: Agents Do Not Remove the Execution Boundary

Expansion is not the same as defensibility in the AI age. Many companies are increasing contract size by selling into labor budgets instead of software budgets. That part works.

What often gets missed is what happens next.

Labor replacement alone is not a moat.

The typical pitch is simple: “We automate what your analysts used to do.”

That can create immediate value. It can justify larger contracts and accelerate adoption. But it does not create a durable advantage.

If the core value is task automation, a competitor with similar models and integrations can replicate it. Model access is not scarce.

Defensibility comes from something else.

The companies that win vertical AI don’t just analyze work. They control execution inside the workflow.

For example, Range has years of client interactions as training data/evals for the agent - that relationship built over time has helped personalize and make it hard to switch as they automate more of the wealth managers’ prior workflows. This is where switching costs begin to form.

Legal AI is a useful example. Consider Ivo or GC AI - drafting memos is not defensible on its own. Embedding into privileged workflows with governance, document access, and compliance constraints creates real friction to replacement so the product gets better with additional use.

Healthcare AI makes the same point more clearly. Model accuracy matters, but integration into EHR systems, compliance with regulatory frameworks, and participation in liability structures are what make these systems difficult to remove. Companies like OpenEvidence and Latent are embedding into the workflow to get patient data, so it’s seamless to close the loop between diagnosis and treatment via finding the path between clinical documentation, payer rules, and pharmacy coordination.

The defensibility is not the model. It’s the execution layer.

AI lowers the cost of building vertical applications. It does not automatically increase switching costs.

If a product remains a recommendation layer without execution authority, it is more likely to trend toward commodity economics.

If you want to understand whether a vertical AI company is defensible, ask:

• Does the product execute work or just recommend it?

• Does it write back into core systems?

• Who owns the liability when it fails?

• Would removing it disrupt operations or just reduce efficiency?

Pricing power follows execution control.

If your product disappeared tomorrow, what would actually break?

→ Part I: Customer Two Is the Only Metric That Matters

→ Part II: Services Aren’t the Sin. Linear Scaling Is.

→ Part III: Automating Work Is Not a Moat

→ Part IV: Agents Do Not Remove the Execution Boundary

Most AI companies look like services businesses early on. That’s not a failure. But it creates a common misconception about what’s actually going wrong.

Services aren’t the problem. Linear scaling is.

Early AI deployments often require embedded engineers, custom integrations, and hands-on execution. That’s rational.

The problem is not that services exist. The problem is when they persist.

For the next customer or existing customer, how hard was it to upsell the existing customer to a new process? Was it just as hard as the first solution? It’s common for consulting companies to be unable to upsell on a bespoke project.

In the beginning, every customer is worth pursuing, but it takes a founder’s serious conviction to turn down revenue, as they know the use case won’t be transversal. There has to be a road map made up.

In a linear model, each new customer requires a proportional amount of effort. More revenue requires more people, more custom work, and more complexity.

In a convergent model, each new customer becomes easier to serve. Early work hardens into reusable systems, and marginal effort declines over time.

Only one of these compounds.

Qualified Health illustrates the transition clearly. Early deployments were deeply embedded and highly customized. Over time, those deployments produced shared ontology models, governance layers, and integration patterns that reduced marginal effort across customers.

The forward-deployed engineers were not the moat. The infrastructure they helped build was.

Ciridae shows the same dynamic from a services-heavy starting point. The company initially relied on human-in-the-loop. Margins expanded only when tooling absorbed the median workflow, and human involvement shifted toward supervision and exception handling.

Automation handles the standard workflows so humans can focus on edge cases.

The signal is not whether services are present. It’s whether they are becoming unnecessary. If delivery headcount scales with revenue, the business is not compounding. It is growing linearly.

If you want to understand whether a company is actually scaling like software, ask:

• Is delivery headcount per dollar of revenue declining?

• Is deployment time compressing across customers?

• Are new customers benefiting from past work?

• Is automation replacing labor in the median workflow?

If revenue grows but delivery scales with it, you don’t have a software business yet.

You have a services business with a software layer.

Are you turning services into infrastructure - or repeating the same work every time?

→ Part I: Customer Two Is the Only Metric That Matters

→ Part II: Services Aren’t the Sin. Linear Scaling Is.

→ Part III: Automating Work Is Not a Moat

→ Part IV: Agents Do Not Remove the Execution Boundary

Most AI startups track the wrong early signal.

They celebrate revenue growth, large enterprise logos, and seven-figure contracts. But those signals can hide whether a business actually compounds.

The earliest real test is simpler:

Customer two should be meaningfully easier to deploy than customer one.

Revenue growth alone does not prove leverage. Things fall as fast as they rise - ask any of the model or coding companies today between releases.

Early AI deployments are often messy. They involve forward-deployed engineers, custom integrations, workflow redesign, and heavy services involvement. None of this is inherently a problem. In high-stakes environments, enterprises aren’t buying intelligence. They’re buying risk removal and operational reliability.

The problem isn’t services. The problem occurs when every deployment stays bespoke.

If each new customer requires new integrations, new workflow branches, and new governance logic, the company is repeating work rather than scaling it.

Compounding begins when early deployments produce reusable primitives. Ultimately, these firms are helping build the blueprint of the next-generation health system for the quote-to-cash process.

Customer one builds the system. Customer two benefits from it.

Palantir is a useful example. Early deployments were deeply embedded and highly customized. Compounding began when ontology models, governance layers, and integration patterns stabilized into shared infrastructure.

The forward-deployed engineers were scaffolding. The compounding asset was the system underneath them.

Scale AI shows a similar pattern. The company initially relied heavily on human-in-the-loop. Margins expanded only when tooling absorbed the median workflow, and human involvement shifted toward supervision and exception handling.

The signal shows up before the financials.

Deployment time compresses. Exception handling becomes standardized. More of each new contract builds on existing infrastructure.

If you want to understand whether a business is actually compounding, ask:

• Is deployment time decreasing from customer one to customer five?

• Is the forward-deployed headcount shrinking per customer?

• Are new deployments reusing infrastructure rather than creating new logic?

• Does new revenue require less engineering effort?

If customer two is not easier than customer one, the business has not started compounding.

Model quality may win the demo. Declining execution effort determines whether the company scales.

→ Part I: Customer Two Is the Only Metric That Matters

→ Part II: Services Aren’t the Sin. Linear Scaling Is.

→ Part III: Automating Work Is Not a Moat

→ Part IV: Agents Do Not Remove the Execution Boundary

During RSA week, I spoke with a few CISOs about how they’re evaluating new AI vendors. The reaction was consistent:

“A lot of this feels like bespoke consulting…
but it’s also how we think we’ll buy software going forward.”

I’ve found myself having the same conversation repeatedly in every category the last few months - from department heads of enterprises like finance to specific verticals like healthcare. AI products look like services early on as they lead with FDE. The question is whether they stay that way.

Revenue growth doesn’t necessarily mean leverage and unwillingness to switch to a new vendor. Large contracts don’t necessarily mean scale. Everyone is testing with real spend and building, and as processes collapse as AI agents do more, it’s not clear whether things will stay when you can rewrite the whole process abstraction in days instead of years. Change management is evolving so fast, and people are adapting in real time.

So the real question is simpler - what actually compounds?

This series is an attempt to answer that from an operator’s perspective.

• When does a services-heavy model become a software business?

• What signals show a company is actually compounding?

• Where does defensibility come from in AI?

As we’ve seen hundreds of pitches of AI services turning into software pitches for every category, I’ll share some observations I’ve found common when working with companies going from early traction to real scale. Hitting these markers isn’t everything, but it does provide a clearer way to evaluate what’s actually happening underneath the surface as people ask whether this is tech-enabled services or a leap toward becoming a software platform.

→ Part I: Customer Two Is the Only Metric That Matters

→ Part II: Services Aren’t the Sin. Linear Scaling Is.

→ Part III: Automating Work Is Not a Moat

→ Part IV: Agents Do Not Remove the Execution Boundary

AI is moving fast. Every week brings another wave of announcements, benchmarks, and breathless takes.

I’m not going to recap the news. There are a million places for that.

What I want to do is slower and more useful: make sense of what’s actually changing — and what it means for how companies get built, funded, and scaled. Because the old playbook has changed. First principles haven’t.

I’ve spent the last decade trying to be the most helpful investor on the cap table — regardless of when we invested or the size of the check. Not the biggest. Not the loudest. The most useful. That orientation shapes everything about how I think about what actually matters at each stage of building.

I’ve been through a reset before. The GFC changed how a generation of companies got built — what capital meant, what sustainable looked like, what founders had to actually figure out on their own. This moment feels different in speed. But some of the underlying dynamics rhyme.

I’m intellectually curious, probably too glass-half-empty, and genuinely trying to make sense of this alongside the founders I work with. These are loosely held opinions - I’m hoping you’ll push back.

The rounds happening right now are dizzying. Valuations that would have been unthinkable two years ago are getting done weekly. Some of it is real. Some of it is kingmaking. Making the distinction is the whole game - and most of the frameworks people use were built for a different era.

The old growth benchmarks — triple triple double — were a proxy for something real: repeatable, leveraged distribution. Companies are now blowing past those numbers. The AI adoption cycle is faster and steeper than anything SaaS produced. But the thing those benchmarks were actually measuring? Harder to find than ever.

A few things I keep seeing:

→ Revenue can scale before leverage appears
→ Large contracts can mask fragile economics
→ Strong retention can look like PMF when you’re actually running a human-powered services firm financed with the wrong capital

The challenge isn’t reacting faster to headlines. It’s understanding what’s actually changing underneath the numbers.

Where I can, I’ll bring data, not to prove a point, but to pressure-test one. And I’ll share observations from the field: what’s getting funded, what the bar actually looks like right now, and where I think the market is pattern-matching to the wrong era.

Signals by Simon are short posts, loosely held opinions and honest questions. Not predictions — I’m not trying to will anything into existence.

The first series starts with a simple question:

AI is labor, so what does that software company actually look like?

Some AI startups can hit millions in revenue within weeks. But fast growth doesn’t mean you’ll last. When anyone can access the same models and push new products in hours, the real edge comes from what compounds — how deeply you’re built into customers’ workflows, learning fast, and surviving even when the platforms move in.

In AI, product launches have become extinction events.

A decade ago, Apple’s WWDC and Google I/O inspired developers to build on top of platforms. Today, when OpenAI hosts Dev Day or Anthropic releases a new model, entire product categories can disappear overnight — the platforms are now building over startups.

Founders joke that “surviving a model release is the new product-market fit.” But speed often hides a deeper challenge: endurance. Speed is abundant; safety, margins, and differentiation are scarce.

The broader debate around AI moats, from distribution to momentum, has been explored widely. A few investors have summed it up well:

“Going from one to fifteen to one hundred is the bar now.” (GC)

“Momentum only matters if it compounds.” (a16z)

Model access is commoditized, infrastructure is cheap, and distribution moves at the speed of a tweet. Growth is no longer the miracle it once was. Endurance is. Others have noted similar shifts as software becomes cheaper and easier to build, but the deeper question now is what actually compounds as the underlying tools converge.

For founders, the goalpost has shifted from signing users to becoming part of how they work. The real moat is when customers depend on you so deeply that “the work stops when you stop.” That’s the kind of value that lasts.

So what actually builds endurance? After watching hundreds of AI launches rise and fade, four patterns keep showing up. Below, we’ll explore this in two parts:

• Part I: The pillars that turn distribution into something durable

• Part II: What still defends in the AI world

  ---

Part I. The Core Pillars of Endurance in AI

Pillar 1: Profit Pools & Buying Moments

AI adoption sticks fastest where pain and budget already exist. Distribution can amplify demand, but it can’t create it.

In healthcare, tools like Abridge and Nabla scale because they turn saved time into billable hours. In legal, Harvey and Ivo ease repetitive drafting and research, protecting margins. In small-business services, voice tools like Avoca and Netic turn missed calls into instant revenue.

These products work because they sell into existing budgets and show ROI fast. When margins are thin, adoption fades after the demo. Founders should sell into funded pain first and create new budgets later. Profit pools, not hype, anchor retention because budgets renew while attention does not.

Pillar 2: Embedded Execution

Software is easier to build than ever. Real staying power comes from being built into the work itself.

Palantir’s enterprise model set the early example. Its forward-deployed engineers made software indispensable by living inside customer workflows. Most AI startups now follow that path, pairing automation with people in the loop to guarantee outcomes and then automating what repeats.

This model works because it replaces or amplifies human effort, creating fast ROI and habits that stick. It’s expensive at first, but it builds credibility before it builds margins. It only scales in high-margin industries or where labor is costly to replace. Smart teams treat it as a bridge to find repeatable workflows they can later automate.

Durability depends on two traits:

1. Integration: you’re part of the daily workflow

2. Adaptability: you keep working as models, data, and infrastructure evolve.

The most resilient AI products evolve with their dependencies instead of breaking under them.

Pillar 3: Full-Stack Learning & Vertical Ownership

Tesla proved the advantage of full-stack integration, where data and control reinforce each other. That same logic drives today’s vertical AI operators.

Companies like Crosby in legal and Range in wealth management act as AI-native operators, rather than tool vendors. For example, Range’s RAI agent turns proprietary client data into a feedback loop that sharpens accuracy and expands margins. In the physical world, industrial and defense operators like Hadrian and Mariana apply similar full-stack logic to manufacturing and mining.

Full-stack models thrive where workflows are closed-loop, repeatable, and high-margin, including healthcare, legal, wealth, and advanced manufacturing. They fail in low-margin sectors where the cost of depth crushes unit economics. Few teams can afford full-stack early because it demands control of data, delivery, and demand. Founders should go vertical only where they can own feedback loops and margins together.

By contrast, open-ended domains such as accounting, HR, or enterprise sales have so far resisted full automation. Human nuance, trust-building, and timing make them asynchronous and difficult to train on. These are the failure modes for autonomous AI. In such spaces, thin horizontal tools that augment rather than replace people may prove to be more durable. The moat comes from amplifying human work and capturing the feedback it generates.

Pillar 4: Platform Gravity

The next generation of durable AI companies will live inside workflows the platforms can’t easily touch. The idea of platform compression — where foundational providers absorb features built on top of them — has been discussed broadly, but its speed in the AI era is unprecedented.

When model providers launch native agents or features, entire markets can shrink overnight, similar to how AWS bundled startup features during the cloud era. The survivors, such as Datadog and Atlassian, embedded deeply in operations or spanned ecosystems.

Regulation and compliance increasingly define those survival boundaries. In healthcare, finance, and government, validation such as HIPAA, SOC 2, FINRA, or FDA creates friction that platforms are slow to address. Compliance becomes a moat when it embeds trust instead of paperwork. It slows convergence and buys time for specialized players to learn within protected workflows.

The same compression now affects voice and video. Even large providers face pressure as model makers move into those surfaces. ElevenLabs, for example, has reportedly surpassed $200 million in recurring revenue, but model-native voice and video will test even scaled incumbents. Defensibility still depends on owning workflows and data, not brand.

We’re already seeing this in motion: OpenAI’s healthcare collaborations and Anthropic’s experiments in financial reasoning show how quickly foundation models climb the stack once traction appears. AI Infrastructure vendors face the same reality: they must prove value through ecosystems or apps that demonstrate why their stack matters.

Platform gravity is inevitable, but partnership often beats replacement. The compression cycle now happens in months, not years. Survival depends on being valuable enough that platforms prefer to work with you, not over you, and on building trust so deep that ripping you out breaks compliance or production.

Endurance in AI doesn’t come from outrunning the platforms. It comes from out-learning them, out-integrating them, and building trust where they can’t move as fast. Speed gets you noticed. Endurance is what compounds. Founders who master both will define the next decade of AI.

Part II. What Still Defends in the AI World

Defensibility begins where the demo ends. Many AI startups confuse early traction with true product–market fit. Each failure mode stems from mistaking fast adoption for embedded (lasting) value.

Skepticisms & Failure Modes

• Ephemeral revenue: pilots look like growth but vanish at renewal. Real revenue lasts through multi-year contracts, seat expansion, and workflow dependence that survives model updates. The faster it comes, the faster it can go, and founders feel that volatility first.

• Synthetic stickiness: integrations that appear deep but are API-thin.

• Middleware mirages: connectors and orchestrators get absorbed by the platforms beneath them.

• Fake data moats: access without feedback loops offers no defense.

• Speed illusion: momentum helps, but learning velocity, how fast outcomes improve, matters more.

The line between pilots and production is blurring. Deals are signed at the pace of experiments from the last cycle, and usage can look strong even when switching costs stay low. In usage we trust, but trust is fickle when migration is easy. The only real proof of staying power is recurring usage that proves its own necessity before renewal.

Learning velocity is now the clearest signal of defensibility. Gains in accuracy, latency, or reduced human oversight predict retention better than revenue growth alone. As models converge in quality, the edge shifts from who builds faster to who learns more per dollar of inference. True defensibility shows up in behavior, not bookings — when customers only switch for a ten-times-better alternative and near-zero friction.

Redefining the Moat

Traditional moats were built on code, brand, or distribution. In AI, defensibility works like a system that compounds through feedback, embedded execution, and trust.

Think of it as two axes: depth × feedback-loop strength.

• Depth measures how essential you are to a workflow.

• Feedback strength measures how much smarter your product gets with use.

The deepest, fastest-learning systems compound; static ones fade as soon as the next model arrives.

Infrastructure and application layers compound differently. Infrastructure moats form through ecosystems, developer trust, and performance or cost advantages. Application moats form through workflow dependence and user learning loops. The best operators know which layer they occupy and build feedback systems that fit it.

Most teams need to pick one axis first: go deep, then learn fast — or learn fast, then deepen. Go deep too soon, and you overfit. Stay shallow too long, and you commoditize.

Atlassian embedded itself so tightly in developer workflows that it became the coordination layer for teams. Datadog did the same in infrastructure, turning observability data into a compounding insight flywheel. Both began as utilities and became platforms. AI companies will follow the same path, from helpful to indispensable, one feedback loop at a time.

For founders and investors alike, the signals of that transition are clear:

• Retention curves flatten above roughly 120% net revenue retention.

• Usage per seat and automation depth rise over time.

• Pilots convert to multi-year renewals.

• Customers start building around the product, not just with it. That’s when switching costs become culture.

The Investor’s Defensibility Audit

1. Depth: Does removing it stop work?

2. Learning: Does it get smarter with every use?

3. Trust: Does it meet regulatory or operational thresholds others can’t?

4. Resilience: Does it keep growing even after the platform moves in?

Defensibility in AI is less about walls and more about gravity, how much value orbits your product even when competitors catch up. For founders, defensibility is not about resisting change but channeling it, building loops that learn faster than the market shifts. For operators, the challenge is embedding those loops where data and dollars converge. Even for investors, the frameworks for AI defensibility are still evolving as models and infrastructure converge.

A moat is not a feature or a model; it is the system that keeps improving when you stop looking. Assume the platform will build your feature or your model. The real test is whether removing you breaks the workflow. Distribution gives you airflow. Execution and learning turn it into lasting enterprise value.