So rather than pontificating about it, I decided to gather some data myself.

Since early August I’ve been pulling daily usage metrics from OpenRouter, which connects developers to about 400 models across more than 80 different providers and tracks aggregate token consumption per model every day. It offers a unique glimpse into genuine revealed preference at scale: not what buyers claim they want, nor what the foundation labs market, but which specific models are actually carrying the workloads.

I was testing a specific idea, one that predates today’s AI wars considerably. One of Clayton Christensen’s lesser-cited observations is that disruption doesn’t begin when a cheap competitor arrives, but when the market leader starts supplying more performance than mainstream customers can absorb. He called it overshooting — the point at which further improvement stops being something buyers will pay for, and competition shifts to price, convenience, or something else entirely.

Answering this question right now is harder than it looks.

What the data does say

Measured on input token rates, roughly 80% of all routed volume is flowing to models priced under $1 per million. Options at $5 or more account for 3.3%.

The most expensive model carrying any meaningful traffic maxed out around $10 per million; premium options listed at fifteen times that saw virtually zero usage.

This largely aligns with what most industry observers, including me, have been suggesting. Low-cost open-weight alternatives aren’t a niche phenomenon here. They’re becoming the dominant choice.

Total volume also grew about 68% over five weeks, from 74 trillion to 124 trillion tokens per week. Whatever else is uncertain, overall demand for inference is not the problem.

Two ways to read the same bill

Now look at that identical traffic priced on output tokens instead of input.

The picture changes materially. On input pricing, 80% of volume sits below $1 per million and 3.3% above $5. Price the very same tokens on output rates and only 57% is under $1, while 10.5% is above $5 — 3X as much.

This matters because output is where the real spend happens. Output tokens carry a premium almost everywhere, and reasoning-heavy models produce vastly more of them, with extended thinking chains consuming multiples of the original input length.

It’s important to be careful here, because I initially thought I had a stronger signal. Tracking both measures weekly, the output-priced share under $1 fell from 63% to 45% while the input-priced share rose — a clean divergence suggesting output was actively repricing upward. But on closer review, that didn’t hold. Across the full window there were 50 increases in output pricing against 65 decreases, and per model, first price to last, 9 went up, 9 went down and 3 were flat. Every one-way increase came paired with a cut elsewhere in the same vendor’s lineup. Yes, the models carrying volume in September do cost more on output than the ones they displaced — but that’s portfolio composition, not a market repricing.

Three things that fooled me

That wasn’t the only thing I got wrong, and the failures are more instructive than the findings.

The disappearing premium tier. The $5-and-up tier at first looked like it was collapsing: share down from 6.1% to 2.4%, and absolute volume down 34% while the market grew 68%. A premium tier shrinking in absolute terms inside a market expanding by two-thirds is almost a textbook picture of performance oversupply.

Except it wasn’t happening. On 18 August one model vendor cut a single model from $5 per million to $2.50, then to $2 four days later. That change alone moved 1.64 trillion tokens a week out of the band — more than the entire 1.53 trillion decline I’d measured. Hold that model’s price at its starting level and the tier’s volume goes up 2%. No buyer’s behaviour changed, rather a vendor changed their pricing and everything adjusted.

Models going to zero. Several expensive models stopped appearing entirely, which read as demand evaporating. But OpenRouter publishes the top 50 models by volume plus one aggregated row for everything else. Falling out of the top 50 doesn’t mean going to zero — it means going into the “everything else” bucket, currently about 6% of all tokens. Four models exited during my window carrying 1.07 trillion weekly tokens, and their successors arrived with almost exactly the same volume, 1.09 trillion. In other words, this is a generational handover, not a collapse.

Prices that move without anyone repricing. Catalog prices are the top provider’s price, and most high-volume models are served by several hosts at different rates. When routing shifts between them the recorded price changes even though no actual decisions took place. One popular model flipped across the $0.10 threshold seven separate times in forty days, each crossing moving 8–14% of that day’s tokens across a tier line. One provider alone accounted for 32 price increases and 42 decreases, flipping back and forth within days.

The high degree of change in such a relatively short period should make anyone cautious about the token-price charts being circulated by both bull and bear pundits.

So is the frontier overshooting?

It’s still an open question that’s not - yet - answerable from consumption data.

Overshooting is a claim about the gap between capability supplied and capability required. What this data measures is capability chosen. The tokenmaxxing era was thankfully short, but it remains true that nobody gets fired for buying the expensive option, at least not at this stage of the market. Corporate inertia pushes revealed preference toward overpaying in ways that can’t be easily discerned from the outside.

What I can say is that the clear majority of real workloads are running well below the top of the market, and have been for at least as long as this observation period. But the move toward cheaper options is coming out of the middle of the price range rather than out of the top, which looks more like ordinary price competition than a softening of demand. And volume in the most expensive tier held roughly flat while the market grew two-thirds.

Answering the important question means measuring what specific jobs actually require. That means taking real business workloads — field extraction, classification, query generation — and running each down the full model ladder to find the cheapest option that gets the job done, against a threshold fixed and published before the results are in. There’s undoubtedly significant use of more powerful models than necessary going on in many organizations, but buyers are getting smarter and vendors have every incentive to adjust their pricing to keep them on board.

Why this is worth watching

It’s worth recalling, as I did in my last post, that two of the most significant AI acquisitions this year targeted infrastructure operating below the frontier rather than the frontier labs themselves. Stripe acquired OpenRouter for over $7 billion to secure the routing layer; Nvidia reportedly committed $12.9 billion for Hugging Face, the open-weight repository hub. Neither bought a foundation model company.

If raw capability were the only scarce asset, value would concentrate at the frontier. Where it’s actually accruing is in the layers that help buyers decide how much capability to purchase, and from whom.

Disruption will come to this market eventually, as it does to all markets. Whether it’s arriving now is a measurement problem, and the current measurements aren’t good enough, especially with highly fluid - and in many cases subsidized - pricing. We’ll keep collecting data, and will update when there’s enough history to say more about direction rather than level.

Data source: OpenRouter (openrouter.ai/rankings), through 13 September 2026. Pricing reflects per-1M token rates from the leading provider per model, calculated against that day’s own catalog snapshot and never carried forward. Price tiers are fixed absolute boundaries, never re-cut against the data. Unpriced models and OpenRouter’s aggregated “other” bucket (5.9% of tokens) are tracked independently rather than distributed across tiers. Newly released models miss their launch day, when the catalog snapshot precedes the release — up to 2.4% of tokens on such a day. Token counts come from each provider’s own tokenizer and are not strictly comparable across providers.

I had to chuckle and told him, “Nice to hear that California investors are catching up to Boston.” <ducking as I say that…>

The New Venture Reality

The current venture market is experiencing extreme concentration, with massive capital flowing into mega-rounds for frontier AI labs and data center infrastructure, while tens of thousands of ventures are struggling to survive. Both Mike Troiano and John Landry have recently written relevant pieces about this structural market shift, suggesting that the traditional venture playbook no longer works like it used to for many. Probably most.

The shift isn’t just about risk tolerance or market cycles—it’s driven by a fundamental rewrite of venture economics.

The New Reality: Validation Before Capital

The old playbook was to formulate an idea, raise a (pre)seed round from investors who you convinced to believe in it, use that capital to hire engineers to build a minimum viable product (MVP), and then hit the market to find out if anyone actually wanted to buy it.

Generative AI has shattered that model. These days, vibe coding platforms and intelligent agents have collapsed the cost and time required to build a working prototype down to, effectively, $0. What used to take six months and six figure engineering salaries can now be spun up over a weekend for the cost of some Claude Code tokens.

Since the cost of building has plummeted, the sequencing has inverted. Market validation now comes before capital, not after it. Code is no longer the bottleneck, rather the scarce resource is proof that customers are willing to pay for what you built. In other words, early-stage capital no longer buys existence; it buys speed, distribution, and scale. Putting capital into a business before verifying market demand is akin to pouring gasoline onto an unlit stove. You aren’t accelerating growth; you’re wasting resources - and risking an explosion.

Redefining Traction in an AI-First World

This forces a redefinition of what “traction” actually means at the seed stage. Traction is no longer a shiny functional app, a waitlist full of personal connections, or vanity metrics like app downloads and free tier signups. Traction today means real prospective customers saying, “This solves a problem I care about, and I will pay you to solve it”—and the ability to show transactions flowing through Stripe to prove it.

Startup creation becomes a rigorous customer discovery problem rather than an engineering challenge. Many founders still stumble here, because they are often builders. Building feels safe—it’s controllable, logical, and gratifying. Customer validation, by contrast, is uncomfortable. It requires stepping out from behind the screen, talking to complete strangers, designing unbiased tests, and avoiding the trap of polite, friendly feedback from people who don’t want to hurt your feelings.

Unfortunately more money won’t fix a founder who can’t clear this bar. It’ll just help them build the wrong thing faster.

The Shift in the Classroom

We are seeing this same trend playing out across top academic institutions nationwide. UC Berkeley’s Haas School of Business recently launched its AI Entrepreneurship course, designed as a hands-on accelerator for MBA founders leveraging AI to accelerate build cycles so they can focus heavily on commercial viability, market research, and customer adoption.

Entrepreneurship education must pivot toward teaching structured customer discovery, interview methodology, and quantitative validation techniques. When AI outsources the build step, teaching founders how to validate becomes the main curriculum.

Practical Takeaways for Founders

Those building a venture or preparing to pitch investors need to adjust their execution playbook:

• Don’t raise to build — raise to scale. Treat fundraising as a catalyst for scaling a customer acquisition loop that you have already proven, not as a source of capital to fund R&D.

• Treat your first 90 days as a validation sprint. Resist the urge to dive straight into IDEs or low-code builders during your initial quarter. Spend those 90 days conducting interviews, testing positioning, and getting clear signals on willingness-to-pay before writing code. Step back from the AI harness and go talk to some potential buyers.

• Apply the “10 Strangers” rule. Before you consider reaching out to investors, force yourself to get 10 complete strangers—people outside your network, alumni group, or friend circle—to commit time or money to your solution. If you can’t convince 10 strangers, funding isn’t your blocker; your value proposition is.

Knowing What to Build Was Always the Hard Part

Returning to that founder I met with: his decision to forego fundraising until he had customers isn’t a sign of weakness, but a sign of modern founder maturity.

Building an MVP was never the true hurdle in entrepreneurship; it was simply the most visible and expensive one. Knowing what to build, and verifying that the world actually wants it, was always the real challenge. AI didn’t change that fundamental truth—it just stripped away the last remaining excuses to ignore it.

On that note, I’m back to the classroom myself this week. Fittingly, one of the courses I’m teaching this fall is oriented around Design Thinking, which preaches this discipline. You don’t start by writing code or polishing a finished product in a vacuum; you start with radical empathy for the prospective user, uncover their genuine pain points, and validate that someone actually cares before committing real resources. Looked at through that lens, maybe the venture machine isn’t "broken" after all. Perhaps the era of raising funding for an unvalidated prototype was the historical anomaly, and the market is simply correcting to how things should have worked in the first place.

I intend to keep writing here during the semester, but my teaching load has increased, so posting may be more sporadic. Either way I’m excited to be back, and confident that the insight and motivation I get from working with the next generation of leaders will help sharpen the sword here as well.

Stripe recently agreed to acquire OpenRouter for more than $7 billion, roughly 5X the company’s Series B valuation from ~90 days earlier. OpenRouter routes requests across 400+ models and 80+ providers, optimizing for task complexity, price, speed, and reliability. Stripe co-founder Patrick Collison has described tokens as the central currency for companies building with AI, supporting the strategic value he clearly sees.

Days later, Nvidia agreed to buy Hugging Face for roughly $12.9 billion - the largest acquisition in the company’s history. Hugging Face is where open-weight models get published, discovered, and forked.

Think about it this way - Stripe bought the toll booth that helps developers avoid paying frontier rates, while Nvidia picked up the hub distributing the models that make that avoidance possible. That’s $20B bypassing the frontier labs entirely and going straight downstream.

The bear case is real, but the framing is off

The concerns about concentration are understandable. Microsoft is booking ~$35B in AI revenue, but a reported $24B of that comes directly from OpenAI. Strip that out, and the rest of the market contributes single digits against more than $250B in capex. Roughly half a trillion in projected cloud AI revenue across Google, Amazon, and Microsoft traces back to just two customers: OpenAI and Anthropic. The chip layer is even more concentrated: one customer accounts for over 15% of Nvidia’s revenue, three drive nearly half, and five represent roughly 70%.

The conventional bear case is simple: the entire AI complex—and by extension the equity market floating on it—rests on the shoulders of a handful of fragile, cash-burning companies.

Noted AI skeptic Ed Zitron laid out his view of this with old friend Josh Brown and Michael Batnick on The Compound and Friends last week: hyperscalers aren’t just funding AI startups with equity; they are offloading the underlying infrastructure risk into private debt markets.

Gary Marcus is wondering whether it will all fall apart.

There’s certainly systemic risk here. When hyperscalers are stashing GPU depreciation inside private credit SPVs, a drop in token volume doesn’t just hit venture investors, it hits debt covenants as well - risking public sector pensions and other parties who may not realize their exposure. But treating an infrastructure financing crunch as a collapse in operational demand misreads what buyers are actually doing.

There are three AI markets, not one

Part of understanding what’s happening is recognizing that there isn’t a single “AI market”. That in reality, there are at least 3 very different segments, which are actively diverging.

1. Frontier AI: High stakes, costly tokens and novel applications. This is where Anthropic and OpenAI play - and certainly don’t count out SpaceX. This segment commands the vast majority of the media’s attention, but over time it will become an increasingly small slice of total unit volume.

2. Consumer & SMB AI: Pure distribution wins here. Nobody checks benchmark leaderboards to pick their personal assistant - at least not in the long run. Google, Apple and Microsoft win on OS defaults and ubiquitous presence. OpenAI’s early lead has all but evaporated, as has their interest.

3. Everyday Enterprise AI: This is ultimately the largest segment, at least in terms of utilization if not in terms of spend. It’s also the least exciting for the media, as in many ways it’s the evolution of what’s been with us for years. Deploying models to summarize customer calls, route support tickets, and extract line item detail from documents. This is the world of repetitive, high-volume and cheap work, where nobody cares what model runs behind the curtain, as long as it doesn’t hallucinate and costs as little as it takes to get the job done. It may not even run in a data center, increasingly it won’t.

Success in one tier does not translate to the others. Consequently, compressed margins in one tier won’t collapse the entirety.

The Everyday market is where open-weight and on-device models will ultimately dominate. Case in point: I dictated the first draft of this piece while driving - using Google’s Edge Eloquent, a lightweight, open-weight Gemma model running locally on my iPhone without sending a single token to a data center.

Enterprise buyers are (thankfully) past the “tokenmaxxing” phase and will increasing look to compress cost per unit of work, just as model choices proliferate. That will certainly constrain growth and margins for the frontier labs, but it is unequivocally bullish for enterprise buyers and the long-term utility of AI.

Big tech is voting with its org chart

Speaking of Google, they framed their recent leadership changes, with Demis Hassabis stepping into the chairman role and Jeff Dean departing, as an acceleration of frontier research. I believe the truth is simpler: Google is focusing their investment where they can get an all-but-guaranteed return - hyperscale infrastructure and ubiquitous distribution. Serving every model—open, custom, frontier, and internal—while pushing Gemini as the default workflow assistant doesn’t require winning every benchmark on the frontier. The WSJ just published a useful piece on how Gemini is helpful for “non-nerds”.

Microsoft is executing a similar playbook. While GPT is available to handle complex multi-step reasoning, everyday commodity workloads are routed to their internal MAI family (reasoning, coding, voice, and transcription), with MAI-Code-1-Flash becoming default in GitHub Copilot.

Amazon never even entered the pure frontier race, Microsoft has quietly hedged it, and Google is optimizing around distribution and infrastructure. None of them need to state publicly that they are no longer investing billions at the frontier—their roadmaps and org charts speak for themselves.

Jevons, not collapse

Cheaper inference does not shrink total spend; it expands total workload volume faster than unit prices deflate. That is Jevons Paradox. When cost-per-token falls by an order of magnitude, enterprises don’t reduce their AI budgets—they embed models into millions of background workflows they previously couldn’t justify running. We are unlikely to see dark data centers, though aggressive unit repricing across older hardware is inevitable.

In other words, the market isn’t shrinking: it’s reshaping. Workloads move down a tier, and margins accrue to whoever controls routing, compute, and distribution. That is exactly the space that Stripe and Nvidia just invested ~$20 billion to secure.

Nvidia’s strategy extends well beyond the Hugging Face sticker price. By developing Nemotron as open source and backing the Nemotron Coalition (Mistral, Perplexity, Thinking Machines, Black Forest Labs), Nvidia ensures that regardless of which open model wins, the workload runs on CUDA. Buying Hugging Face simply secures the storefront.

What we are seeing across enterprise deployments is classic Christensen overshooting: the gap between what top-tier models can do and what day-to-day business workflows actually require has widened into a chasm. In disruptive innovation theory, this isn’t failure; it is the exact inflection point where the market fragments and a lower-cost tier absorbs the volume.

The frontier labs will endure, because high-stakes, zero-error enterprise tasks will always justify premium compute. But they will no longer define the entire ecosystem. They never did.

Speaking of old friends, Thomas Otter just wrote a thoughtful piece referencing my last one. In Where is this epoch’s Hammer and Champy?, he explores what happens when the cost of generating software collapses to zero—a good reminder that while software generation is becoming abundant, the hard constraint shifts to organizational change management, knowing what not to build, and aligning technology with human systems.

Work in Progress by Thomas Otter

Where is this epoch’s Hammer and Champy?

It was a sunny Johannesburg spring day in 1993. The jacarandas were in bloom. I took the afternoon off work to drive from Midrand to a specialist business and techni…

Read more

18 days ago · 1 like · 1 comment · Thomas Otter

This is what makes the Substack writer community so valuable: when technology and unit economics shift overnight, having a network of thoughtful practitioners challenging assumptions and sharpening each others’ saws helps us all make sense of rapidly evolving markets and their real-world implications.

In other words, that’s a hint—I’d love to hear your thoughts.

Leave a comment

Here’s how we got here.

Rusty Beemers on 128

I arrived in the Boston area and the software industry in the early ‘90s. For those not old enough to remember, this was right at the end of the microcomputer era. I can still recall my first boss telling me about all the rusted-out BMWs cruising Route 128, the result of ongoing layoffs from Wang, Prime and DEC. Ironically, we sold software to those companies’ customers.

The customers got the memo, but took some time to act on it, as they still had their VAXes, their PDP-11s, and their IBM mainframes plugged in and running mission-critical work. Nobody was unplugging what worked, not quickly at least. That’s the first lesson, and it repeats: the vendor economics die years before the installed base does.

Where it got interesting—and fun for me—was when that started to change.

From Big Iron to Enterprise Software

My next role landed me in the middle of the rise of “client-server.” Some have only heard that term in a legacy context, but the term itself doesn’t matter much. What it enabled is what’s important: a new platform for computing, and for software running on a network, rather than a single machine.

What client-server ultimately enabled was the thing we ended up calling Enterprise Software. SAP. Oracle. PeopleSoft. Siebel Systems. A long list of companies, some still with us, others long forgotten.

Alongside the vendors came the consultants. I particularly remember Michael Hammer, whose Reengineering the Corporation provided the intellectual scaffolding for the whole era. “A Manifesto for Business Revolution” no less!

The Reengineering pitch, in a nutshell: we have this new computing platform, and it can help enterprises (of course) re-engineer themselves. Which really meant changing how they operate by automating their processes.

Adapt the Business to the Software

Here’s what that looked like in practice. SAP — the leading example in the early days — shipped standard modules. Finance. HR. Warehouse management. Supply chain. Support tickets. You bought it, ran it on top of your client-server network, and it standardized those processes for you.

Which was genuinely appealing, because how do big companies get big? They scale. How do they scale? They standardize. Enterprise software was a standardization machine sold to organizations whose entire growth model depended on standardizing, and who never before had the ability to standardize on an enterprise-wide basis.

But notice the order of operations: it was not about adapting the software to your business. It was about adapting your business to the software.

Yes, there was flexibility at the margins. SAP famously had ABAP, which let you customize your processes to a degree. But you were building on top of what the software gave you. The application’s model of how a company works was the starting condition, and your business was the variable.

Importantly, this was a massive improvement over what came before. Wildly popular, and in some cases wildly successful. In other infamous cases, catastrophic. But successful or not, the enterprise software decision was a standardization revolution, and just about every large customer signed it.

The Departmental Revolt

In the late ‘90s, that starts cracking with the rise of business use of Netscape and the internet, which suddenly promised the opportunity for users to run anything as long as it ran in a browser. We weren’t calling it “SaaS” yet — the precursor was Application Service Providers (or ASPs). Web-hosted software, delivered via the ubiquitous web browser.

What ASPs enabled were compartmentalized, departmental solutions. There was real demand for that, because inside every large organization you would have senior leadership looking at enterprise-wide processes — but you also have the xVPs of Sales, Marketing, Supply Chain, HR, etc…. And to those people, a monolithic ERP system was too much. Too heavy, too slow, too far from their actual problem. Not to mention far too expensive for the CFO, especially considering the armies of high-priced consultants who needed to fly in every week to install it, customize it, and support it.

There was an accessibility problem too. Early enterprise software and mobility didn’t mix. The iPhone didn’t exist yet, since phones made phone calls but they didn’t run apps, at least not apps robust enough to be useful. Users had laptops, but syncing from a remote environment was a miserable experience of moving the bed away from the wall in the hotel to sync over a slow dial-up modem before you could get any real work done.

Eventually ASP evolved into the SaaS we know today: innovators like Salesforce disrupting Siebel and Workday doing the same to PeopleSoft demonstrated that enterprise capabilities could really be delivered in a browser, and widespread broadband made the SaaS promise real.

SaaS offered a more compartmentalized approach to process management. Rather than a company-wide re-engineering exercise — which the monolithic ERP systems required — you could bring on a departmental solution to run sales, or marketing, or the warehouse. Less comprehensive. Far more flexible. Built for the xVP, not the C-suite.

The downside was that it steered us away from true enterprise-wide process integration. APIs showed up to patch that, and mostly did.

The Covid Adaptation and the ZIRP Blowout

All of this went on for quite some time - bandwidth and browsers kept improving on the front end, and APIs and web services on the back end. Integration got easier - if not perfect - and the world settled into a rhythm. Investors loved it and predictable metrics like ARR and NDR became metronomes of the SaaS economy, and those rusty Beemers were much harder to find.

Then came the pandemic, and everything kicked into hyperdrive, when two things happened at once.

First, every enterprise suddenly had to work remotely, and the existing SaaS stack didn’t always fit how work now suddenly had to happen. New solutions emerged to fill the gap - Zoom being the most obvious example: built for the suddenly remote environment everyone was abruptly living in.

Second, governments unleashed a ton of free money, as interest rates went to zero and loan programs like PPP recharged corporate balance sheets. The SaaS market, in turn, went a little nuts. From late 2020 into early 2022 there was an absolute blowout of new SaaS providers funded on cheap capital. What we actually got was an over-proliferation of me-too next-gen products. And the old stuff was all still running.

Which brings us to present day.

Where We Are and Where We’re Going

Today’s market has a vast over-proliferation of SaaS, with the market working off an oversupply of me-too point solutions, and CFOs and CIOs partnering to consolidate their tech stacks. I did a survey about this with CIO.com in 2023 and the customer feedback couldn’t have been more clear.

There’s a hierarchy to SaaS. At the foundation, the core systems of record, from the SAPs and Salesforces, still run not just departments but entire business units. Attached to those SORs sits a sprawl of single-function point solutions, what I’ve always called the solutions in the App Store.

Every one of those SaaS solutions, from the largest ERP down to the smallest point tool, is sold on the same fundamental bargain that was struck in the early ‘90s. The vendor built one model of the process. You adapt your business to it. You pay per seat for the privilege, though there’s some experimentation around the edges with usage-based and value-based pricing.

Until, suddenly, the advent of ChatGPT and the generative AI wave. What started as chat interfaces quickly revealed where the real leverage was: generating code. Almost overnight, the core economic reality that dictated thirty years of IT strategy began to collapse. Led by Anthropic’s Claude Code, Cursor, Lovable, Replit and others, suddenly enterprise attention is turning from an era of buying standard software to one of generating bespoke systems.

The Reversal

For thirty-plus years, software was expensive to build and cheap to copy.

That single economic constraint dictated the entire structure of the enterprise software market. It’s why vendors amortized one rigid, opinionated process model across thousands of customers. It’s why enterprise buyers absorbed the friction between that generic model and how their business actually ran. The mismatch was a massive, recurring operational tax—but because building from scratch was impossible, everyone paid it and called it “best practices.”

Now, the cost of generating software is rapidly collapsing to zero. The entire bargain inverts.

This isn’t an incremental efficiency gain for the SaaS playbook; it’s the elimination of the premise SaaS was built on. Incumbents cannot defend against this by dropping seat prices, because customers were never buying the seat. They were buying someone else’s process—out of necessity.

When you can direct autonomous agents to spin up, integrate, and continuously modify systems tailored precisely to your operating model, the concept of a multi-tenant application suite starts looking as dated as a client-server rack in an on-prem data center.

The vendor economics always die long before the installed base does. The legacy SaaS stack won’t vanish overnight, but the mandate has already shifted: the thirty-year compromise is over, and the leverage has returned to the enterprise.

Transitions this fundamental only come around every few decades. The new playbook is barely being written, and it is going to be thrilling to watch.

Leave a comment

My position is simple: as a reader, I don’t particularly care.

What I care about is whether it moves me. Whether it teaches me something. Whether it was worth the time I gave it.

Reading has purposes, and they aren’t the same

I read for two fundamental reasons. To be informed, and to be entertained. I tend to do too much of the former and not enough of the latter, but that’s another topic.

But in informed mode? I find AI to be enormously useful. I consume a lot of AI-authored content and a great deal more AI-assisted content. As long as it delivers what I came for, the provenance is a minor detail that I really don’t spend any time thinking about.

No doubt the entertainment case is different - the human touch means a lot there, and I have a hard time believing that AI-driven fiction would move me the way a great writer would. Although I’m open to being surprised there too - I’ve got a few friends using Suno to make music that is quite entertaining - and meaningful.

My classroom experience

I’ve been teaching at Northeastern for a few years now, and AI-assisted writing is of course as hot a topic in a university as it is anywhere.

I’m lucky that I teach in a business curriculum. My students (mostly) understand the value of these tools, and they’re clear-eyed about the fact that future employers will almost certainly care far more about their ability to leverage AI effectively than their ability to compose elegant prose. That’s not to say writing skills don’t matter — but the ratio has shifted dramatically. I’m not certain how I’d approach this if I were teaching a topic like English, but that’s also not my problem to solve.

When I started teaching, I thought I could spot AI-written work. I didn’t necessarily penalize students for using AI for aforementioned reasons, but it factored into how I read their submissions. As I say all the time in the classroom, my job is to make sure they understand the material, not that ChatGPT or Claude does.

My confidence, however, collapsed quickly. Within a few short semesters it became all but impossible to tell, and I’d suggest I have a better eye for it than most.

For a short while, I tried AI detectors, but I found them wildly unreliable. Certainly not something I’d feel comfortable assigning a grade against, and I stopped treating their output as evidence of anything.

Watermarking changes things - somewhat

As you may have heard, just recently Anthropic began embedding an invisible watermark in text generated by Claude. This is applied at the model level, so it’s present regardless of which product the text came out of, and it survives copy-and-paste. Files get signed provenance metadata under the C2PA standard. This is driven largely by the EU AI Act’s transparency requirements, which nearly 200 companies have now signed onto, including OpenAI, Google, Meta and Microsoft.

I don’t see this as a bad thing. I use Claude heavily, and yes that includes here. The research we publish at Differential Factor is substantially enabled by Claude, and we’ve never hidden that. Just the opposite, it’s part of the value proposition.

If regulators and readers want a machine-readable signal, they should have one.

But notice what watermarking actually solves. It makes detection reliable — which is a real improvement over the guesswork we have now. It answers the how, but it doesn’t make the question matter any more than it did before.

As detection is improving, my curiosity about the label is declining. As a reader, I care about the effect, not the machinery.

If the answers are yes, the provenance is a footnote.

Also worth mentioning: Substack itself shipped Pangram-powered scanning last month, after CEO Chris Best published a piece arguing against what he calls “Claudefishing.” His concern is a mismatch between what a reader expects and what they get — attention invested in something with no human thought behind it. That’s certainly a valid concern for the CEO of this platform to address, especially since writers are his customers.

But I’d frame the problem a bit differently. The problem isn’t that no human wrote it. The problem is that no human thought about it. Those aren’t the same thing.

Take this post. I dictated the entirety — the position, the classroom story, the objection to detectors — and then handed the transcript to Claude to recompose, and to Gemini to help refine. More and more that’s how I’m doing my “writing”. The thinking is 100% mine. The sentences get - help. And because the watermark is applied at the point of generation, it can’t tell those two things apart. It sees text that came out of a model and marks it accordingly, which is technically accurate, but not particularly useful, at least if you’re interested in understanding what I think.

I don’t write for a living

Finally, I fully understand that this is disruptive, which forces change, and people tend to hate change, particularly when it impacts their livelihood. None of this should indicate that I don’t feel for those who make their living writing. Their fear and frustration are fully legitimate, and I’m not attempting to wave that away.

But there’s a real productivity story to be told: that AI doesn’t just speed up writing, it can also make it better. I 100% believe that to be true, and I hope this post is some small evidence of that.

Much of this current debate is transitional anxiety. In a few years, or perhaps sooner, consuming AI-generated content will be so ordinary that the question stops being interesting. We’re just living through the loudest part of this transition. Not too far down the road, I doubt many will care.

And great human writers will survive - and thrive. The best will be even more productive. Of that, I’m quite certain. It’s not a matter of if, it’s when.

I’d love to hear where you land on this. Even if you disagree. Especially if you do.

Leave a comment

My model called that final in June, before a ball was ever kicked.

It called it again the week of the final. Same answer both times: Spain over Argentina.

No massive database. No real-time odds tracking. Just a simple set of rules running in memory.

The experiment

Like most I’ve been building at Differential Factor, the World Cup model started as an experiment. I wanted to test some new predictive modeling workflows - and Anthropic just happened to release their new Claude Fable model around the same time2. As an avid soccer fan, the World Cup offered the perfect data-rich playground.

Going in, I assumed a frontier model would help me design something complex and powerful - a system that could ingest match stats, account for tactical coaching shifts, and pick non-obvious upsets.

That’s not what happened.

Instead, what Fable helped me build was small. Very small. It is printed below, and it runs in a rounding error’s worth of memory.

{
 "params": {
  "version": "v1.0.0",
  "K": 50,
  "HOST_ADV": 100,
  "ALT_PER_1000M": 35,
  "HEAT_PER_10C_OVER_28": 12,
  "REST_PER_DAY": 8,
  "REST_CAP_DAYS": 3,
  "GD_SCALE": 250,
  "GD_CLAMP": 2.5,
  "TOTAL_GOALS": 2.7,
  "MIN_LAMBDA": 0.15,
  "GOAL_GRID": 9
 },
 "version": {
  "id": 1,
  "version": "v1.0.0",
  "created_at": "2026-06-11 13:58:55",
  "params_json": "{\"version\": \"v1.0.0\", \"K\": 50, \"HOST_ADV\": 100, \"ALT_PER_1000M\": 35, \"HEAT_PER_10C_OVER_28\": 12, \"REST_PER_DAY\": 8, \"REST_CAP_DAYS\": 3, \"GD_SCALE\": 250, \"GD_CLAMP\": 2.5, \"TOTAL_GOALS\": 2.7, \"MIN_LAMBDA\": 0.15, \"GOAL_GRID\": 9}",
  "change_note": "Initial public model. Covariate coefficients are declared assumptions, not estimates."
 }
}

And it nailed the outcome that mattered most.

Where it lost

To be clear, this isn’t a victory lap.

Across all 104 matches, the betting markets beat my model handily—especially in the group stage. A tiny set of rules can’t account for a team resting starters once they’ve already qualified, lineup leaks, sudden injuries, or squad drama. Betting markets, of course, are experts at capturing real-time news and subtle coaching shifts.

On aggregate, the markets won. But on the single question that mattered most, the simple model got it right.

Small vs the supercomputer

All summer, Sports Illustrated ran projections from the Opta supercomputer. Massive infrastructure, continuous data pipelines, updated after every match.

Look at how Opta’s predictions shifted over time. France kept climbing while Spain drifted down. Of course, France didn’t even make the final.

To be fair, my model wasn’t entirely immune to matchday noise either. As the live tracking chart shows, a surge in France’s late-stage metrics temporarily bumped them ahead in our simulations around July 9th.

But here’s a crucial difference: because my model relied on a tight, bounded set of core rules, it quickly recalibrated back to its foundational premise. It called Spain over Argentina in June before the tournament started, and held that final stance in July.

Opta’s supercomputer, by comparison, had no way to pull itself back. Because it ingested five weeks of live match data—shot charts, expected goals (xG), match results—it continuously refined itself away from the right answer.

More compute. More data. Continuous tuning. All driving a massive system further in the wrong direction.

Two brief caveats before I draw enterprise parallels:

• Probabilities aren’t predictions. Opta outputs probability distributions, not binary picks. Giving France a 34% chance isn’t “wrong” when Spain wins. The issue isn’t Opta’s math - it’s the media narrative framing it as a “supercomputer prophecy.” Those same misplaced expectations are too often forced on enterprise AI - and AI in general.

• n=1. Calling one final correctly is a nice story, but not statistical proof. Over ten tournaments, Opta’s engine would likely win on average.

The Job to be Done

The Job to be Done here wasn’t “outperform a supercomputer across 104 individual matches.” The job was “predict the tournament winner.” For that specific outcome, a tiny heuristic model was entirely sufficient.

Would routing this problem through a massive, multi-billion-parameter frontier model have shaved a few percentage points off our variance? Probably. Would it have changed the actual business outcome? Not at all.

Yet in most enterprise AI discussions, this distinction gets completely buried under executive panic.

When a board asks, “What’s our AI strategy?”, the instinct is to provision the biggest, most capable model available. Nobody wants to look like they’re cheaping out or falling behind in the race to become “AI-native.” But provisioning an expensive frontier model without scoping the task is like hiring a team of PhDs to sort incoming mail.

Or like driving a Ferrari to the grocery store, as I wrote about a few days before the tournament.

So what is the job, actually?

Most enterprise tasks aren’t open-ended reasoning puzzles. They are bounded, specifiable, and repetitive:

• Classify an inquiry

• Extract three fields from an invoice

• Route a support ticket

• Summarize a call

Well-defined jobs don’t need frontier models. They need right-sized models. Focused models that run quickly, cheaply, and reliably—and often run perfectly well on a local device or edge node.

Where the frontier earns its money

Saying “small models are often enough” isn’t the same as saying “frontier models are a waste.”

The frontier is where synthesis and ideation lives. It’s built for ambiguous problems, novel framing and cross-domain reasoning.

Look at what actually happened in this experiment: I used a frontier model once, to act as the architect. I asked it to help me figure out how to frame the problem, combine team skills ratings (Elo) with expected goal distributions (Poisson) and then build a Monte Carlo simulator to run thousands of bracket outcomes.

Once Fable helped write those core rules, its work was done. I no longer needed a multi-billion-parameter LLM to run daily simulations or track match-by-match Brier scores, which is how we ranked against the betting markets. I just needed the tiny, lightweight script it helped me create.

That is the core pattern. Frontier to think. Small to run.

The mistake isn’t using frontier models - it’s overusing them as runtime workhorses. Deploying a massive, high-cost model for a daily execution job means paying a steep and unnecessary compute tax on every inference. Right now, that’s exactly what too many organizations are doing.

My commute as an example

I drive a Tesla Model 3, which is a terrific commuter car for my trips to and from Northeastern University, where I teach. The local model in my car is very good at one thing: keeping me on the road. It’s fast, local, and hyper-optimized for its specific environment.

But as I sometimes tell my classes once I arrive—that same model would be completely useless at filing my taxes.

My World Cup model would, of course, be useless for both. But for predicting the tournament, it matched the scope of the problem effectively.

Matching the tool to the task used to be standard engineering discipline. Somehow, in the AI gold rush, enterprise leadership has lost that instinct entirely. The default setting has become reaching for the largest, most expensive model available—regardless of what it’s actually being asked to do.

Why this is a capital story

This isn’t just an operational detail—it’s a macro capital story.

Today’s AI infrastructure spending surge relies heavily on a single implicit assumption: universal frontier dependency. The belief that every enterprise workflow, agent, and query must route through the largest, most expensive cloud models available.

That assumption underpins massive capital allocations across venture portfolios, hyperscaler projections, and corporate IT budgets. But a tiny heuristic model correctly predicting the World Cup isn’t just a fun math story—it’s a direct counter-argument.

The rapid rise of open-weight models is forcing enterprise leaders to rethink their deployment strategies—driven primarily by pure cost pressure. While open-weight doesn’t automatically solve the engineering complexity of hosting, fine-tuning, and maintenance (a topic so deep it deserves its own upcoming post), the price gap between proprietary API calls and self-hosted models is hitting CFO radar screens hard. C-suites are being forced to stop throwing raw compute at every problem and start asking what the job actually requires.

Match the model to the job. Most of the time, the job is much smaller than you think.

Traditional research was built for a world where market boundaries were static, vendor categories stayed in neat boxes, and buyers had time to read 40-page PDFs to justify a decision.

Today, tech does not stay in boxes. When a leadership team meets to address a critical issue, the debate is not “Which of these five SaaS tools in the upper right quadrant should we buy?” It is instead a competition for internal resources: Do we buy software, build a custom fix in-house, change our operational process, or reallocate that budget to a different initiative entirely?

That is why we built Competition for Capital (CfC), a first-principles approach to market intelligence that looks at how real decision-makers allocate time, budget, and attention.

A First Principles Approach (Not Another 2X2 Matrix)

Traditional analyst firms force the market into rigid coverage areas because human analysts can only track a dozen companies at a time before their capacity breaks down. They draw artificial boxes around categories to make them fit into a 2x2 grid or similar static scorecard. They do this not because that is how buyers think, but because that is how the firms package their subscriptions.

CfC flips that model entirely.

Check out Competition for Capital (CfC)

Instead of forcing your problem into a pre-defined map, CfC starts with the question you are actually trying to answer. Market boundaries are not fixed, instead they overlap, intersect, and evolve in real time based on buyer demand. CfC maps the option space dynamically, giving a view that evolves as quickly as the market does.

How to Use CfC

Most buyer decisions fail long before a contract is signed because the different stakeholders around the table are working from completely different perspectives.

A CEO is thinking about risk and resource allocation. A CFO is looking at ROI and payback periods. A sales leader is tracking competitor positioning, and a VP of Product is trying to figure out what to build next.

CfC is built around a single, shared source of market truth, tailored to the specific question you need to answer:

• For Buyers: Stop asking “Which of these 20 SaaS apps should I buy to fix churn?” Start asking “How should we approach our churn problem from first principles?” CfC provides guidance whether the right answer is buying a point solution, tweaking a process, or reallocating budget elsewhere.

• For Sellers & Founders: Get an unvarnished view of where your solution actually sits in the market, including the non-traditional alternatives, internal builds, and adjacent tools your buyers are evaluating against you behind closed doors.

• For Investors, Recruiters & Job Seekers: Evaluate a company’s true competitive moat, market trajectory, and positioning before committing capital - or your career.

Why “Capital”?

When we say “capital,” we do not just mean cash. Money is often the easiest resource to measure, but it is rarely the rarest.

Every strategic decision is a trade-off of time, engineering resources, executive focus, and team capacity. When you choose one path, you are implicitly burning capital in all those other forms. CfC measures the entire trade-off, not just the subscription price on an invoice.

Built on AI, Grounded in Experience

Yes, there is serious AI under the hood here, utilizing frontier models including Anthropic's Claude for analysis and reasoning, and Google’s Gemini for embeddings powering retrieval.

Anyone can type a quick prompt into ChatGPT and get a generic summary. CfC goes much further. We have combined sophisticated reasoning with a custom, high-dimensional knowledge engine that maps relationships across thousands of market signals.

More importantly, the system is architected around the methodology I have used over 30 years of evaluating tech markets. It synthesizes market dynamics through an analyst’s framework. As more users query the system, the underlying knowledge layer gets progressively smarter - for everyone.

Pricing: Simple & Grounded

Running high-reasoning AI models and maintaining dynamic knowledge bases incurs real infrastructure costs. Because Differential Factor is completely independent and self-funded, we have tied platform access directly to Substack. This keeps the tool accessible without putting it behind a traditional 5+ figure corporate paywall.

Everyone with a work email gets their first run free — subscribers can use any address.

From there, your usage is tied to your Substack tier:

• Free Subscribers: 1 run every 30 days

• Monthly Paid Subscribers: Up to 5 runs every 30 days

• Annual Subscribers: Up to 10 runs every 30 days

• Founding Members: Up to 25 runs every 30 days

(Note: Setting up recurring quarterly market updates counts toward these totals).

No sales pitch, no hidden upsells, and your contact info will never be rented or sold.

Subscribe now

Try It Out & Send Feedback

The best way to understand Competition for Capital is to put your own actual business question through it.

Run your first CfC analysis here

Once you’ve run a report, drop a note through our feedback form. Tell us what worked, what surprised you, and where the analysis missed the mark. Your feedback is what will shape the next iteration of the platform.

Thanks for reading, and I look forward to hearing your thoughts.

Local farmers were locked out of credit by tightening bank lending limits. So, they worked with an agtech firm called Cowmed to put their herd on the blockchain. Each cow wears an AI-powered “smart collar” that tracks its health, behavior, and location in real time. That data feeds an encrypted digital identity into a credit agreement registered on Brazil’s B3 exchange.

The system prevents a farmer from pledging the same cow to two different lenders—the classic fraud risk in any collateral-based loan—and even lets them swap in a live cow if one dies.

The pilot raised nearly $20,000. Cowmed already tracks about 100,000 cows worth over $395 million, and it expects up to a fifth of that herd to eventually finance itself this way. That unlocks roughly $77.6 million in new agricultural credit.

It’s a small and whimsical story. But I believe it’s a clear leading indicator of something much bigger.

Finance is overdue for disruption

I’m among those who have long believed that tradional finance is one of the industries most exposed to classic Christensen Disruptive Innovation. The structural reason is easy to overlook: there is no intrinsic reason that government-backed “money” needs to be the mechanism of exchange between two parties.

Money has earned its historic role by making exchange simpler and safer, backed by sovereign trust at scale. It’s effectively invisible when the system works—when credit flows freely, when banks want your business, and when the cost of capital is reasonable.

The Paraná farmers show what happens when the traditional system fails. Local banks tightened lending limits on small agricultural producers, effectively walking away from a segment of real, productive economic activity.

That is exactly the type of environment where disruption gains a foothold. It doesn’t start by competing with the incumbent on its home turf. Instead, it serves an underserved need that the incumbent has explicitly decided isn’t worth serving.

To a bank relationship manager, a tokenized cow is a worse “loan” in almost every dimension that matters to them. So they ignore it. Until gradually, then suddenly, it’s competing with them. And disrupting their business.

The Interchange layer is next

Once tokenized value transfer proves itself in underserved corners—agricultural credit today, remittances and small-business working capital tomorrow—the pressure moves toward the core plumbing of everyday commerce.

As every small business owner knows, the legacy Card Networks (Visa, Mastercard, American Express and others) extract 2 - 3% on the overwhelming majority of consumer transactions in the developed world. That toll, generally referred to as “Interchange”, was historically justified by the trust, fraud protection, and settlement guarantees they provided.

But as tokenized rails accumulate their own track record on those exact dimensions—verifiable identity, fraud prevention, and real-time settlement—the rationale for routing routine exchange through a legacy interchange tollbooth gets thinner every year.

What’s in your agent’s wallet?

To be clear, “crypto” has gotten a bad name over the years. High-profile collapses like Mt. Gox and FTX, and more recent political shenanigans, have provided plenty of high profile reasons for consumers to be cautious if not outright skeptical. In human-to-human commerce, trust is heavily influenced by these rare but sensationalized headlines.

But this shift is bigger than consumer crypto, because commerce is increasingly going to become agentic.

Humans have their reasons to prefer a familiar card network: force of habit, brand trust, and of course, those highly-valued reward points. AI agents transacting on behalf of a business or individual have none of those attachments. They will optimize ruthlessly for cost, speed, and verifiable settlement. A network that can cryptographically prove an asset’s value in real time (the way Cowmed’s collars prove a cow is alive and unpledged) is a natural fit for a counterparty that has no use for a loyalty card in its wallet.

Trust is the barrier, and that barrier is falling

The single force holding back tokenized finance hasn’t been technology. It’s trust.

Governments earned their monopoly on money primarily through force, and partly through centuries of demonstrating that the promise on the note would be honored. Tokenized systems have had to build that trust from scratch, transaction by transaction, in the wake of some very public failures.

What this little cow story shows is how quickly trust can now be manufactured at the point of the asset itself, rather than borrowed from a central authority. The collar, not a regulator or a bank examiner, is what makes the cow trustworthy collateral.

That’s a genuinely new pattern: verifiable trust engineered directly into the asset, accessible to anyone who can read the data, rather than trust conferred top-down by an institution. As that pattern gets replicated across more asset classes—inventory, invoices, real estate, and eventually financial instruments themselves—the last real advantage the incumbent system holds keeps shrinking.

Tokenized transactions raise massive questions around visibility and tax compliance, often rendering economic activity effectively invisible to government authorities. That regulatory tug-of-war is a deep topic for future posts, but it’s a flavor of disruption that goes straight to the foundations of our political, economic and legal systems.

None of this means banks and card networks disappear next quarter, or even next decade. Disruption theory has never been about speed; it’s about direction. But the direction here looks very consistent.

As you can see, the model puts Spain ahead at 53.5% to Argentina’s 46.5%. Certainly not a landslide, but the numbers favor La Roja. Interestingly, Spain being the favorite over Argentina is exactly where we started on day one.

Spain got here by looking like the better team in practically every round, capped off by a dominant 2-0 semifinal win over France. Argentina, on the other hand, barely survived. They needed extra time to get past a massive underdog in Cape Verde during the round of 32, and then had to mount a thrilling comeback win against England in the semis.

Both pre-tournament favorites made it to the final, but only one of them made it look easy.

How the model did

We (my friend Claude and me) built this model to be intentionally simple: no scouting network, no proprietary data feed, no black box. The goal was never to out-think the pros1. It was to see how far a simple and honest model could get on its own.

The final numbers look better than I expected. Through the group stage, our simple model called winners correctly 60% of the time, while the market hit 67%. By the semifinals, that gap narrowed to nearly a third. For a bare-bones project tracking a multi-million dollar market backed by real money and massive data feeds, staying in the conversation feels like a win.

But I have to put a pretty big asterisk on that scorecard: the main reason the model looked so smart in the back half is that the world behaved exactly the way it was supposed to - because the favorites kept winning. The four teams almost every serious pre-tournament projection — including ours — flagged as the real contenders were Argentina, Spain, France, and England. Those are exactly the four teams that reached the semifinals. A final between the two pre-tournament co-favorites is about as predictable an outcome as a 48-team tournament can produce.

This tournament had a stretch of upsets early on in the group stage, and that’s exactly where the model hiccuped. A simple mathematical model relies heavily on historical class and baseline ratings. What it can't factor in is human chaos, such as a massive underdog like Cape Verde playing with absolutely nothing to lose, or a team resting its starters because they've already locked up a spot in the knockouts as the USMNT did against Turkiye. Once the tournament settled down in the knockout stage and the contenders sorted themselves out, so did the model’s job. Worth remembering before crediting the method for what the field mostly did on its own.

One more thing before we close it out

Even in the home stretch, we ran into a stubborn bug on the UI side that took a full day and several wrong guesses to track down. It didn’t touch the actual predictions, but it was a healthy dose of humility and a reminder of a basic truth when you’re building things: a system can look perfectly polished and confident on the outside while being quietly broken on the inside.

I wrote about this dynamic earlier in the tournament in Confidently Wrong —it’s the classic engineering trap where a bug behaves exactly like a feature until it suddenly doesn’t. You don’t fix those moments by spinning up a more complex theory or cleverer reasoning. You fix it by stopping, throwing out your assumptions, and demanding the system show you exactly what it’s doing under the hood.

What’s next

Regardless of what happens tomorrow, the full model, every prediction, and the scoreboard stay public at wc2026.differentialfactor.com. I’m also planning to use this experience in the classroom at Northeastern this fall as a teaching moment or two.

I’m considering whether I want to build another model for the upcoming Premier League season, but I may just choose to enjoy the games instead. #COYG

The Information recently published a piece on this exact trend, describing how a number of organizations are using low-code tools such as Claude Code, Lovable, Replit and Cursor to move away from enterprise SaaS platforms like Salesforce, HubSpot, and ServiceNow.

There’s an active discussion in the comments, which I’m part of, and one commenter made a point I fully agree with: it would make complete sense for companies like Anthropic and OpenAI to offer a baseline CRM framework. There’s a ton of horizontal functionality in the CRM space that’s already fairly commoditized, not to mention it’s a massive TAM for Anthropic and OpenAI.

The comments also include the usual counterarguments, of course:

Have you factored in the cost of support?

How will you handle maintenance and upgrades?

Are you considering how you’ll servicing hundreds of users?

Those are all fair concerns - but there’s no denying that trend is real, and companies are actively running build-versus-buy calculations - and more than a few are starting to choose “build”. One company saved $100K, another $40K, and even Sanofi (not exactly an SMB) is saving “millions”.

Build vs Buy Calculator

Keep in mind that, even for buyers who ultimately decide not to go down the low-code or no-code path (at least not yet), simply having that option gives them significant leverage in negotiations with their SaaS vendors. That alone is already reshaping the market.

This is textbook Disruptive Innovation. It starts with a lower-end, seemingly inferior solution that incumbents dismiss as “not good enough.”

Until, suddenly, it is.

As Thierry Henry said, just magical.

I took a look at the model this morning in a good mood: it’s been performing well. And then I looked at today’s fixtures and noticed something was off. It had Canada playing Paraguay and France playing Morocco. The pairings were scrambled.

World Cup 2026 Living Model

I dug in, and I found two bugs (well, Claude did). Worth detailing both, especially the second which is the most interesting one.

The first was mundane: bad bracket data in the seed file had a couple of the knockout fixtures wired to the wrong slots. Annoying, quick to fix, verified against the official feed, corrected in the live database.

The second one had been hiding in plain sight since the tournament started. The model, by design, doesn’t treat draws as a weakness: it predicts them, and a tie doesn’t count against a team’s strength rating. That’s intentional. But knockout games can’t end in a draw; they go to a shootout. And a penalty shootout isn’t really a measure of which team is better: it’s close to a coin flip, which is exactly why I didn’t want it feeding the strength ratings.

The problem was how the old code handled that. When a knockout match came back level after regulation, it silently assumed the away team advanced. Not because that’s true (it obviously isn’t) but because I hadn’t explicitly told it what to do, so it guessed. And it had gotten lucky three times in a row, quietly picking right, which is the most dangerous way for a bug to behave.

I fixed it so the model now does the honest thing: when a knockout ends level, it stops and flags the match for a human to enter the real result, rather than inventing an answer. A shootout isn’t a strength signal, so the model shouldn’t pretend to read one.

Here’s why I’m bothering to write this up.

I said on the webinar with Athena Solutions last week that the thing you have to watch for with AI isn’t that it’s wrong: it’s that it can be confidently wrong.

This little bug was a perfect small example. The model wasn’t uncertain about today’s matchups. It didn’t hedge. It served up “Canada vs Paraguay” with the same crisp assurance it uses for everything else. I caught it because the fixtures were checkable against reality.

The uncomfortable question is how many confidently-stated outputs aren’t that easy to check - and catch. The discipline isn’t teaching the machine to doubt itself. It’s remembering that you have to.

A day later and the wrong results would have propagated down the bracket. Instead it’s fixed, deployed, and documented.

One more thing: what if Cape Verde had actually won?

Since I had the model open, I ran a counterfactual. The betting market has been ahead of my model all tournament. I wrote about that earlier; a transparent strength model gives up ground to a market that prices in everything it can’t see. But the gap’s been narrowing in the knockouts, and last night’s match is a good window into why, and how fragile that is.

My model and the market both had Argentina as a heavy favorite, but the model was more confident: 83.7% Argentina versus the market’s 79.3%. Argentina won, so on that match the model actually beat the market, one of the nights it closed the gap.

But flip it. Had Cape Verde pulled off the massive upset, the model wouldn’t just have lost; it would have lost to the market, and by more than it gained last night. On the Argentina win, the model beats the market by about 0.025 on that match (Brier score, lower is better). On a Cape Verde win, it trails by about 0.10. A swing of roughly 0.13, against the model.

That’s the mechanism behind the narrowing gap, and its catch. The model has been clawing back ground on the market partly by betting harder on Argentina than the market does. Every time the favorite wins, the model gains. But that’s not a smarter model; it’s a more concentrated one. The same conviction closing the gap now is exactly what would blow it back open the night the favorite falls. And in a single-elimination bracket, that night ends Argentina’s tournament and the model’s momentum in the same instant.

Last night, that night was one goal away.

People Stuff

Why Would They Buy You?

A founder showed me a product recently that solved a real problem for a real customer. Clean design, working software, paying users. Two years ago I’d have spent most of my diligence on the technology — whether the team could actually build what they were describing, whether the CTO had the depth to scale it. That question barely matters anymore. The pr…

Read more

3 months ago · 7 likes · 4 comments · Mike Troiano

Potential acquirers are fundamentally valuing a mix of three assets - the “3 T’s”.

1. Talent

2. Technology

3. Traction

Every valuation calculation is a three-legged stool. Kick one out and it tips. Kick out two and you’re left with a post.

The Talent leg has been getting wobbly for some time. Silicon Valley figured out the acqui-hire - if the team is the primary asset, it’s quite easy to strip it out and just buy it directly. From an operator’s perspective, this can be a decent outcome - assuming you’re part of the core team - but from an investor’s perspective it’s almost always bad.

Now AI is coming hard for the Technology leg. When anyone can vibe-code a working product in a weekend, the technology itself stops being scarce. The head start you thought you had — gone. That’s the real shift Mike is pointing at, and it’s the one many founders are understandably reluctant to price in.

So what’s left standing? Traction. Paying customers and revenue you can bank. In many respects, that’s how it always should have been.

A story that rhymes

About a dozen years ago I was advising the CEO of a social media startup, where over time I got deeply enough involved that I effectively became part of the founding team. We had a prototype that (mostly) worked. We had early “design partner” customers - not yet really paying us but helping us refine. In other words, we didn’t have much Traction yet, but we had Technology, and we had Talent (of course…).

We were also highly dependent on the large social media providers for our data - one in particular. This was early enough in the social media era that the big social networks were more focused on growing their audiences than on providing analytics around their data - so they made that data available to their “ecosystem” - fed through a small set of approved data partners. Like scores of startups, we were floating our business in that “partner” ecosystem.

One day the CEO called, excited: the big company invited us out to the Bay Area to show them what we were building. My first reaction was a flicker of “wait, why.” Acquisition interest, investment interest, market validation — could be any of those. Could also be none of them.

One of our advisors, watching from further out, told us not to take the meeting without protection. But no NDA was on the table — “just an introductory conversation” — so that was that.

A well-known VC we saw the day before said it was a bad idea. But tickets were already bought. So we went.

The meeting itself was pleasant. Polished. They loved what we’d built, talked about “exploring how we could work together,” and by the time we wrapped it was late so we headed back to the hotel to rest up for some potential investor meetings the next day.

The next morning my phone rang early. Our lead engineer, a few hours ahead in the UK, had already informed the CEO: our data feed had been cut off.

Of course, there was no way prove a direct line from the prior day’s meeting to the cutoff. Nor was there any smoking-gun email. But the timing was not subtle. And it turned out we weren’t special — over the next few weeks and months, nearly every startup in that ecosystem suffered the same fate.

In hindsight, it was closer to a mercy killing than a betrayal. We’d come close to raising a few times, but our early product was never the same thing as a business, and the path forward was foggy at best. I don’t carry regrets. I carry the lesson.

“Partnerships” are about self-interest. Nobody sits across the table from you out of charity — they’re there for their own self-interest, full stop. Walk in with that mindset, and you won’t be surprised by the outcome.

Traction is the only leg that matters

For struggling startups today, that lesson matters more, not less. You may be thinking that AI has handed you new leverage - but it’s at least as likely that it’s taken more away. The technology head start you were counting on is worth less than it used to be, because everyone can build fast now.

Which brings me back to that 3-legged stool. Two legs are going if not gone. The one that’s left — the one that was always the most important one — is your Traction. Traction isn’t just usage, but real customers paying you real money because you solve a real problem for them.

Go find some. And then go find some more. It’s still the only leg that holds weight.

I often say that watching soccer is like watching golf — if you’ve played it, you can appreciate what’s happening; if you haven’t, it can be a slog. I’ve watched a lot of soccer these past few weeks, and this tournament has been the opposite of a slog. It’s been three weeks of the best stuff.

Not to mention that here in New England we’ve had some great side plots. The Tartan Army literally drank Boston dry - which many didn’t think was possible. And then there’s my sentimental favorite Norway (my family heritage) - watching them row their way through the group stage has been a delight.

Also credit to Mauricio Pochettino. I’ve spent a lot of years trying to watch the USMNT, and this is the first time I really feel like I’m actually watching a *team* — not just a collection of athletes. I’ve played - and coached - a lot of soccer, and one thing the game teaches you, over and over, is that it’s a team sport in a way Americans don’t always appreciate. We tend to over-celebrate individual athleticism over the unglamorous stuff — spacing, movement off the ball, eleven players solving a problem together. This USMNT finally looks like a true team. Credit to them and the coach who catalyzed it, however far their run goes.

Anyway, on to the model.

The scoreboard

Three weeks ago we launched a World Cup “living” forecast model1 — a live demonstration of research that commits before the fact instead of explaining itself after. Every prediction locked and cryptographically hashed before kickoff. Methodology fully public. A scorecard that updates after every match, including the ugly ones.

Access the Model

The group stage is done. All 72 matches were predicted, locked, and scored. Here’s how it went.

The model called the right outcome 43 times out of 72 — 60%. The betting market called 48 — 67%. On Brier score, which is the standard way to grade probabilistic predictions, the model came in at 0.540 against the market’s 0.450. Lower is better, so: the model lost. Consistently. It never beat the market over a full phase of the tournament.

That was the expected outcome - and it’s also the point.

The model lost because of draws, and that’s the whole story

20 of 72 group matches ended level. The model — like every model built on team-strength ratings — is bad at draws. They don’t get predicted so much as left over, and when a bunch of tight and cautious opening-round games finish 1-1, the model eats the cost every time.

The market is better at this, because the market isn’t really a model. It’s everyone with money on the line, including people who know *this particular match* is going to be a grind for reasons that never show up in a rating.

Which gets at what the group stage actually demonstrated.

The gap between the model and the market isn’t a failure - it’s a measurement

The model knows exactly one thing about each team: how good they are, updated from results. It does not know who’s injured. It doesn’t know that a team has already qualified and is resting starters (such as the US resting 9 starters against Turkiye). It doesn’t know that another team is eliminated and just playing for pride. It doesn’t - and can’t - read the news.

The betting market, of course, knows all of this, and prices it in.

So the gap between them — that 0.09 on the Brier score — isn’t a mystery. It’s the value of everything the model can’t see. Lineups, motivation, injuries, who needs what from the table. And the only reason we can point at that number and tell you what it represents is that the model is simple enough to be honest. When it’s wrong, it tells you exactly why: it priced a full-strength USA against Turkiye, and Coach Poch (wisely) rested most of the squad. The market is more accurate and can’t tell you a thing about what it knows or why.

Those are two different tools. The market is built to be right. The model is built to be explainable. The space between them is the data I actually care about.

The vibe-coded model vs the “supercomputer”

Sports Illustrated ran Opta’s post-group-stage projections this morning — the “supercomputer” predicting the winner. Opta is the real deal, an industrial forecasting operation with far more going into it than my deliberately blind little teaching model. So it’s worth seeing where a one-channel strength model - vibe-coded in a few hours - lands next to the supercomputer.

At the top? Basically the same answer. Opta’s three standouts are France (18.7%), Argentina (16.3%), and Spain (13.5%). The model’s top three are the same three teams — Argentina (19.1%), Spain (17.4%), France (11.7%) — shuffled in order but unmistakably the same tier, with the same clear gap down to everybody else. Two models built on completely different philosophies agree on who the likely winners are.

Where they split is the interesting part. The model likes Colombia a lot more than Opta does (7.3% vs 3.2%), and prefers Argentina over France. Both of those trace straight back to how the model works: it learns only from results, and Colombia and Argentina had excellent, clean group stages. A pure strength model rewards exactly that. Opta, with more inputs, smooths those swings back toward what it expected before the tournament. Nobody’s been proven right yet. The knockouts decide it — which is the whole point of locking both sets of numbers down before kickoff.

What happens now

The Round of 32 starts this afternoon, and the first knockout predictions are already locked. Two things change from here, and both work against the model: the games get tighter, and the stakes get higher — which is exactly when the market’s edge over a blind strength model is biggest. If the model holds up anyway, that’ll be a pleasant surprise. If it gets taken apart, that’s the cleanest demonstration yet of what a strength-only model can - and can’t - see.

Either way, it’s already on the record.

Access the World Cup 2026 Model

Most companies are rushing into AI, but they’re skipping the most critical step: ensuring their data is actually ready for it. If your data is fragmented or ungoverned, no amount of shiny new AI models will save you from the “garbage in, garbage out” trap.

Register Now

On Thursday June 25, I’m joining forces with Alex Kangoun (CEO, Athena Solutions) to cut through the noise. We are focusing on the practical, unglamorous work required to turn your fragmented systems into real, AI-ready intelligence.

We’ll be breaking down:

• The 5 Dimensions of AI Readiness: Strategy, Data, Process, Technology, and People.

• The Governance Gap: Why data quality is the prerequisite for reliable AI results.

• The 4-Week Roadmap: How to assess your readiness and build a plan that actually scales.

This session is designed for business leaders who want less hype and more execution.

Join us:

• Date: June 25, 2026 | 12:00 PM EDT

• Register: Fix the Data First: AI Readiness & Data Governance

Hope to see you there!

To mark the opening whistle, we are running a live, public test of an AI-native predictive framework. The timing was partially dictated by the rapidly-evolving landscape: with Anthropic’s Fable model becoming available just this week, this offered an incredible opportunity to test what is quickly being acknowledged as the most powerful model on the market. The catch? We had to “vibe code” this entire “living model” architecture from the ground up in a wildly compressed time window right before kickoff.

Access the Live Pipeline Tracker

Fun meta-note: Fable built the actual production engine; Sonnet & Gemini helped draft this post to get it out before kickoff.

The Methodology: No Retrospective Hedges

Most sports analytics models operate with a comfortable degree of revisionist history. This living model rejects that. It commits before kickoff, not after.

Every single prediction is locked, given a cryptographic fingerprint, and hashed before a ball is kicked. When the final whistle blows, it is scored publicly against the live betting market—including when it loses.

Each hash represents an absolute snapshot of the model’s environment at that second: team ratings, situational covariates, and the model version. Post-match, anyone can pull the endpoint (GET /api/match/1), recompute the data, and verify that nothing was touched.

Here are the hard commitments for today’s opening slate:

📊 The Day 1 Board & Projections

Match 1: Mexico vs. South Africa

• Venue: Estadio Azteca (Mexico City)

• Kickoff: 3:00 PM ET

• The Model’s Lean: Mexico 79.2% · Draw 15.2% · South Africa 5.6%

• Verification Receipt: 81a79e929df2a66d2813adb10d64767c7556fa50fa8bb583156a860f34bdaa75

Match 2: South Korea vs. Czech Republic

• Venue: Estadio Akron (Guadalajara)

• Kickoff: 10:00 PM ET

• The Model’s Lean: South Korea 38.8% · Draw 25.8% · Czech Republic 35.4%

• Verification Receipt: 785e248b5b50069ef3e301d2b1334a9c623fa77f05a34339ef867a220da3f261

Behind the Numbers: Two Crucial Analytical Notes

If you are looking at these percentages and comparing them to standard market consensus, two distinct features of Fable’s reasoning stand out:

1. Why Mexico’s Line is Massive (And it’s not the altitude)

A 79.2% win probability for El Tri looks incredibly heavy on paper, but the internal logic is fascinating. Many analysts assume playing at the iconic Estadio Azteca automatically triggers a steep physical penalty against opponents due to the extreme altitude.

However, the model’s altitude covariate is currently flat. Because this is Day 1 of the tournament and both squads are arriving fresh out of structured training camps, the immediate physical toll of the elevation is minimized. The altitude variable only begins to scale and index heavily starting with each team’s second match.

Instead, this line is driven purely by a stark fundamental rating gap combined with a massive baseline host-nation advantage.

2. The Korea/Czech Coin-Flip

In contrast, tonight’s late-night matchup in Guadalajara (which I am unlikely to be awake for thanks to OG Anunoby) is an absolute dead heat. The model places South Korea at 38.8% and the Czech Republic at 35.4%.

The model isn’t being timid or trying to hedge its downside here; the mathematics simply dictate a coin-flip. The raw Elo ratings sit at 1758 vs. 1740. When two mid-tier sides with almost identical structural metrics meet on neutral ground, the algorithm refuses to manufacture an opinion where a clear edge does not exist.

🏆 Tournament Outrights: Who Lifts the Trophy?

While Day 1 gives us our first verifiable data points, the living model isn’t just looking at daily matchups. It is constantly recalculating the macro probabilities of who actually wins the tournament. If you check the live pipeline at wc2026.differentialfactor.com, you’ll see the outright futures board.

Right now, the model has planted its flag on two heavyweights:

• Spain 🇪🇸 (The Strong Favorite): Fable’s underlying logic heavily rewards Spain’s structural metrics. The algorithm highly values teams that control possession and limit high-variance transitions, viewing that systemic control as the ultimate tournament-survival trait.

• Argentina 🇦🇷 (The Close Second): The defending champions follow right behind. They retain a massive baseline Elo rating, and the causal brain clearly respects their proven knockout-stage pedigree.

Because this is a living model, these are not static, fire-and-forget predictions. Every goal, card, and tactical shift over the next month will dynamically ripple through the framework, adjusting each team’s odds in real-time.

What’s Next?

We’re tracking the ROI, calibration, and structural integrity of this fully vibe-coded pipeline publicly throughout the tournament. If Fable crushes the market, the prompt architecture holds. If it goes down in flames, we document the hallucinations.

You can audit the live trackers, pull the raw data streams, and verify the hashes directly at wc2026.differentialfactor.com.

Drop a comment below: Are you riding with the machine or fading the model on Day 1?

Leave a comment

Disclaimer: This project is an autonomous software and data science experiment conducted for research and entertainment purposes. It does not constitute financial or sports betting advice.

Of course, Microsoft is playing its own strategic game of catch-up here - they’re not entirely altruistic in making this argument. But that doesn’t make them wrong.

Everyone’s Driving a Ferrari to the Grocery Store

Let me use an analogy that I think captures what’s actually happening in AI right now.

The financial markets are subsidizing Ferraris — and giving everyone the ability to drive one. There’s nothing wrong with Ferraris - I’d love to have one. They look spectacular, they’re undeniably thrilling to drive, and they impress your neighbors. But are they really what you need to run errands? To pick up your kids from school? Is burning that much horsepower, that much fuel, that much money the right call for every single trip you take?

Of course not.

The same question applies to AI models — and right now, most people aren’t asking it.

The AI industry is spending enormous time and energy on every lab’s latest frontier release. The hype around Anthropic’s Claude Mythos — a model that, as of this writing, isn’t yet publicly available — is a perfect illustration of this dynamic. Frontier models are absolutely remarkable and incredibly powerful. But they are also breathtakingly expensive, and a growing share of those costs are being obscured by investors who are still funding market share over margins, and by model providers who are actively incentivizing adoption over efficiency.

That cannot last forever. And when it stops, the reckoning will be uncomfortable for those who never developed the habit of asking: is this the right model for this task?

The Hidden Tab

Here’s what the current moment looks like beneath the surface.

Yes, per-token costs are falling — that’s real. But the unit economics argument misses the bigger picture. Frontier reasoning models consume orders of magnitude more tokens per task than their lighter counterparts. As agentic workflows scale, and as organizations pile on more complex, multi-step AI processes, volume outpaces the per-token price drop. The bill goes up even as each line item looks cheaper. It’s the classic “we’ll make it up in volume” fallacy playing out in real-time.

The anecdotes are starting to pile up. It’s been well-publicized that Uber’s CTO recently disclosed that the company burned through its entire 2026 AI coding tools budget in four months — after internal leaderboards aggressively incentivized adoption, without anyone tracking what, exactly, they were optimizing for. That’s not a story about AI failing. It’s a story about an organization that belatedly developed the discipline to ask whether the tool it was reaching for was the right one for the job. Other organizations such as ServiceNow, Walmart and even Microsoft and Amazon themselves have been dialing back their encouragement to their teams to max out AI usage.

These are early warning signals. Most organizations — and most individuals — are not yet reading them.

Christensen Had a Word for This

Those of you familiar with disruption theory will recognize the dynamic immediately.

Christensen called this dynamic overshooting — the moment when a technology’s raw capabilities outpace what the mainstream market actually requires. When that happens, chasing the absolute frontier yields diminishing returns for users, while quietly compounding the costs. Overshooting creates the classic conditions for disruption from below: a “good enough” alternative, priced for accessibility, eats the market from the bottom up while the incumbent keeps racing toward a frontier most customers don’t require.

We are watching this happen in real time in AI. Frontier models are extraordinary — genuinely so. But for many, arguably most, business tasks, they are already well past “good enough.” The question isn’t whether they’re impressive. The question is whether the incremental capability justifies the incremental cost.

For most tasks, it doesn’t.

Seeing This From the Inside

I run a business. I use AI constantly — it’s foundational to how I work, not a novelty or a tool I reach for occasionally. When you get the bills - even the subsidized ones - it becomes crystal clear that defaulting to the most powerful model for every task is inherently wasteful.

Drafting a routine email? A smaller, faster, cheaper model handles it better than you’d expect — and returns the result in a fraction of the time. Summarizing a document? Same story. Doing deep comparative research across a complex strategic landscape? Now you need the horsepower. That’s where frontier reasoning earns its expense.

The discipline of matching model to task — what I think of as right-sizing your AI — is something I now treat as a core operational practice. It saves money. It saves time. And frankly, it often produces better results, because a well-scoped smaller model frequently outperforms a frontier model given a vague, underpowered prompt.

But most people haven’t developed this habit. “Tokenmaxxing” is dumb - but it’s still the mindset in too many places. The default setting — especially when someone else appears to be picking up the tab — is too often the most powerful option available.

The Reckoning Coming

The subsidies won’t last forever - investor patience for labs burning capital to subsidize market share is not infinite. Granted, with the massive SpaceX, Anthropic, and OpenAI IPOs looming on the horizon, the hype machine will likely drown out talk of capital efficiency for a while longer. But the music will eventually stop.

When pricing starts to finally reflect real costs, organizations that have built workflows around the assumption that frontier performance comes at commodity prices are going to face a hard reset.

The smart move, for individuals and for organizations, is to get ahead of this now. Not by abandoning frontier models — they have a real and important role — but by developing genuine fluency in the question that the industry currently discourages you from asking:

Microsoft is asking it, even if their motives are mixed. CIOs and CFOs are starting to ask it loudly. The Uber story won’t be the last one.

You should be asking it too.

It's not having what you want

It's wanting what you've got

(with apologies to Sheryl Crow)

Now that I’ve had a chance to decompress a bit after an intense and exhilarating Boston Tech Week, I wanted to weigh in on a debate that’s been buzzing in the local ecosystem. Earlier this week, my old friend Doug Levin posted a characteristically un-subtle critique, expressing disappointment that a California-based firm (Andreessen Horowitz/a16z) led the week rather than local Boston leadership.

Lessons from a Startup Life

Shame on You, Boston: You Can’t Sponsor Your Own Tech Week?

Starting today, Boston hosts its first-ever Tech Week, running May 26 through 31, 2026. The target audience is exactly what you’d expect: “the top funds, startups, and companies around the globe” with more than 570 events spread across the city. Boston Dynamics, Wayfair, MIT, MassBio, Anthropic, the Red Sox and Bruins analytics teams. The lineup reads l…

Read more

4 months ago · 7 likes · 6 comments · Doug Levin

While Doug’s post sparked plenty of defensive pushback on LinkedIn, to me, this whole controversy is a tempest in a teapot.

First off, credit where it’s due: we should be thanking a16z (which Doug did!) for putting their massive marketing muscle, visibility, and global wrapper around our city’s ecosystem. A tech week is defined by the community that shows up to fill it—and the Boston community showed up in spades. Local organizers leveraged the momentum perfectly to host some truly memorable events.

Boston’s numerous assets were on full display all week: our people, our companies, the city itself, and, of course, our world-class universities (yes, I’m entirely biased here).

On that academic note, we hosted an inspiring inaugural event at Northeastern University, bringing together and celebrating our rapidly expanding entrepreneurial community. Having the Tech Week banner allowed us to reach far more students, alumni, families, and faculty than we ever could have on our own.

The momentum carried through the rest of the week, too. The venerable MassTLC Unconference, masterfully put on by Sarah Fraim and her team, was a terrific opportunity to catch up with old friends, meet some new ones, and talk through fresh ideas. And to cap it all off, the TBD Angels social event—which was 3X oversubscribed thanks to the heroic efforts of Jessica Angell Carlson—was the perfect, high-energy networking note to close out at least my week.

This regional “Boston vs. Silicon Valley” debate has been going on since I first moved to Boston in the early ‘90s, and frankly, it’s getting a bit silly. Perhaps I’m a little jaded as a native New Yorker, but tech and AI buildout is not a zero-sum regional competition. If it were a pure race, let’s be honest: the Bay Area is so far down the road we can’t even see their taillights.

Instead of squabbling over who gets to promote Boston, we should be focusing on how much there is to celebrate right here. If there is a central sun in the current AI universe, it’s a16z—and it’s to their credit that they are shining a light on what’s happening globally, including right here in our backyard. Our job isn’t to fight for the head of the table; it’s to firmly take our rightful place at it.

It was a phenomenal week. I hope everyone gets a chance to step back, enjoy the sun (both literally and figuratively), and have a great weekend.

I'm, I'm gonna soak up the sun

I'm gonna tell everyone to lighten up

But there is an uncomfortable reality that many organizations are hitting head-on: AI is only as good as the architecture supporting it. When systems are fragmented, the data flowing into your models is inherently flawed. For business leaders, this isn’t just a technical headache; it’s a strategic risk. Moving too fast without a unified, governed data foundation leads to unreliable intelligence, flawed market predictions, and wasted investment.

To address this, I am partnering with Alex Kangoun, CEO of Athena Solutions, for a live webinar: Fix the Data First: A Framework for Turning Fragmented Systems into AI-Ready Intelligence.

Learn More and Register

What We Will Cover:

• The AI Readiness Framework: A look at the five critical dimensions required for successful adoption—Strategy, Data, Process, Technology, and People.

• Building a Trusted Foundation: Why strong data quality and robust governance are absolute prerequisites for reliable, actionable intelligence.

• From Strategy to Execution: How management teams can assess their current readiness in just 3–4 weeks and build a scalable roadmap that evolves alongside changing market dynamics.

Event Details:

• Date: June 25, 2026

• Time: 12:00 PM EST

• Format: Live Zoom Webinar + Q&A

If you are looking to move past the AI hype and build a practical roadmap for scalable implementation, we’d love to have you join us!

Register Here

There is no shortage of opinions on this trend—including our own—but at the moment, there is very little actual data. While the industry is rife with speculation about how easy AI makes it to “Build,” (and counterarguments of the not-so-hidden costs and risks) there is a clear gap in evidence regarding how organizations are actually navigating the “Buy” in the real world.

We’re aiming to help fill that data gap.

Today, we are proud to launch the Build vs. Buy Decision Platform at bvb.differentialfactor.com. We are making this software available for free because we want to help move this discussion from opinions to evidence. With this tool, we have a mechanism to collect the anonymized, behavioral data required to understand how corporate decision-makers are actually weighing these trade-offs right now - and how those decisions will evolve over time.

Check out Build vs Buy Platform

Meet Harry, our Enterprise Software Agent

To analyze this data, we’d also like to introduce you to Harry, our resident Enterprise SaaS agent. As our first agent at Differential Factor, Harry is here to navigate the strategic and financial metrics and decisions involved in making a Build vs. Buy decision—helping teams move beyond anecdotal opinions toward evidence-based, computational judgment. Whether you are evaluating a new purchase or preparing for a high-stakes renewal, Harry helps you weigh the variables that actually influence outcomes: total cost of ownership, speed to value, and the strategic drivers of your business’s actual needs.

Robust Enough to be Useful, Simple Enough to be Easy

BVB is a V1. We have intentionally kept it lean because we know that for a tool to be useful, it must be usable. Our goal was to create a framework robust enough to support a defensible business case, yet general-purpose enough to be used across different categories of SaaS.

A few things to know as you explore the platform:

1. BVB is absolutely free. This is a research initiative intended to support analysis of key enterprise-level decisions.

2. Built for renewals and new purchases. Use it to stress-test your current vendors and help your negotiation before you sign your next multi-year contract.

3. Harry is a trainee. He’s already quite good at analyzing inputs, but over time as he gathers more data he’ll become even better at benchmarketing, analysis and recommendatinos.

4. A Contribution to the Data. By using the tool, you are helping us build the first real dataset on how the “Build vs. Buy” calculation is evolving in the age of agentic AI. Your data will be kept fully anonymous and will not be used for any purpose beyond analysis.

The SaaSpocalypse doesn’t mean the end of software; it means the end of making decisions in the dark. We invite you to put it - and Harry - to work.

Visit bvb.differentialfactor.com and let us know: does Harry’s logic match your vision? We’re looking forward to your feedback and appreciate your participation.

Build vs Buy Decision Tool

“I don’t know what will happen to the group of today’s SaaS incumbents.” Companies that address AI head-on can do better than ever, while others may “lose market value, go bankrupt or completely go bust”, he said.

Dario Amodei, CEO and Co-founder of Anthropic, May 6, 2026

I have been spending a lot of time in the classroom at Northeastern these past few years talking about Clayton Christensen’s Disruptive Innovation theory, but this past semester it really hit home—as it became clear that this dynamic was happening in the world where I spent the majority of my career: Enterprise Software.

Market commentators have dubbed this the “SaaSpocalypse.” After fifteen+ years of "up and to the right” growth, the SaaS industry appears to be hitting a wall. Traditional “per seat per month” pricing is rapidly falling out of favor with buyers, metrics such as ARR, NDR and LTV are wilting, point solution fatigue has led to ongoing vendor consolidation (which I predicted - based on CIO input - in 2023), and valuations are being slashed in both private and public markets.

Some are taking the position that this is a standard market correction. I believe that there’s something much more fundamental taking place: classic disruptive innovation.

The Investor Trap: Confusing “Sticky” with “Safe”

For nearly two decades, the prevailing investment thesis in SaaS has centered on the assumption that these solutions were particularly “sticky.” Investors made massive bets, operating under the assumption that once a company implements and integrates a specific ERP, HRIS, or CRM, the switching costs—the organizational pain of “ripping and replacing”—act as a protective barrier.

But in the Christensen framework, a moat built on the difficulty of departure rather than the delight of the Job to be Done is a fragile one.

For years, SaaS incumbents have focused on sustaining innovations—adding more features and more complexity to justify price increases to their most demanding customers. In doing so, many have “overshot” the needs of the average user, creating a massive opening at the low end of the market.

Enter the “Inferior” Disruptor

Christensen taught us that disruptive innovations often start by looking like toys. They are “inferior” by traditional performance metrics - but offer a new value proposition: they are simpler, cheaper, and more convenient.

Today, that disruptor is Generative AI and vibe-coded solutions—software built not by massive engineering teams over years, but by small teams using LLMs to generate functional code on the fly.

By traditional enterprise standards, these AI-driven agents and lightweight tools are inferior. They lack the twenty+ year history of SOC2 compliance and the massive implementation manuals of a Salesforce or an Oracle. And of course those now-legacy systems are often the systems of record, where the “keys to the kingdom” data and core workflows reside. Thinking of replacing them has been a career-ending risk - until the cost of not doing so becomes even higher.

However the order-of magnitude leap in developer productivity that AI is providing changes the equation - allowing small teams to build highly specialized and customized agents that solve their users’ Jobs to be Done with a fraction of the friction. As long-time enterprise software executive and investor Brett Queener recently noted, we are moving toward a world where value is found not in the tracking of data (the system of record), but in the judgment applied to it. This judgement varies - for competitive reasons - from organization to organization; so this isn’t just a cost issue, it’s a competitive one as well.

The “Headless” Pivot: Delivering Value, Destroying the Seat

The most telling evidence of the SaaSpocalypse isn’t coming from the startups; it’s coming from the incumbents themselves. Take Salesforce’s Headless 360 announcement. By exposing the entire platform as a headless set of APIs and MCP tools, Salesforce is admitting that for many users, the “Job to be Done” no longer requires a browser.

This is the right move for customers—it removes the “UI tax” and allows AI agents to operate directly on the data. But it also fundamentally blows up the traditional SaaS revenue model. You cannot charge $150/month for a “seat” when there is no one sitting in it.

One has to give Salesforce and its peers credit for having the strategic foresight to disrupt themselves—this is the Innovator’s Dilemma in action. They are choosing to cannibalize their own “per-seat” revenue model before an AI-native solution does it for them. Whether they can successfully navigate the revenue chasm on the other side remains to be seen.

It also remains to be seen how committed the industry really is to disrupting itself, as half-measures by some providers limit the usefulness of their “disruptive” initiatives.

Let’s not forget that Kodak invented the digital camera.

Vertical Resilience vs. The LLM Acceleration

Others have suggested that industry-specific Vertical SaaS is a safer place to be. Perhaps. The “Job to be Done” in a specialized field like oncology or heavy construction requires deep, localized context that generic AI may not be as able to grasp as easily as the fairly standardized context of ERP, CRM & HRIS applications.

Even if we believe that vertical solutions have a higher likelihood of remaining sticky in the near term, no one should confuse “safer” with “safe.” The “contextual moat” is evaporating faster than many anticipated because the models themselves are specializing.

Anthropic’s recent release of a specialized legal module for Claude is a massive shot across the bow for vertical software, as are the aforementioned finance models referenced earlier in this piece. If an LLM can be “taught” the nuances of BigLaw or banking in a matter of months, the industry-specific moat that vertical SaaS providers have relied on for years will eventually look more like a puddle.

The Shift from Seats to Outcomes

Thinking back at my years in enterprise software - the “SaaSpocalypse” is sobering but not surprising. We are seeing the beginning of the end of the “SaaS-as-a-Service-Tax” era.

• The Horizontal Risk: Platforms for horizontal functions - Finance, HR, CRM - are ripe to be disrupted because they are the most predictable and the least differentiated. While in the near-term they are the hardest to 'rip and replace,' that's a defensive posture, not a growth strategy. The disruption is currently most active at the edges, where AI agents are devouring the high-margin 'point solutions' that once thrived in the incumbent's app stores. In effect, Salesforce’s headless pivot is a tactical retreat - that sacrifices its own AppExchange ecosystem - and the many point solution vendors in it - than to its core.

• The Vertical Warning: Specialization is no longer a permanent shield. As specialized LLM modules expand, the “contextual moat” is evaporating. The adaptation of elite professional services firms - and the birth of well-funded giants like OpenAI’s The Deployment Company and Anthropic’s $1.5B partnership with Blackstone and Goldman Sachs - proves that AI-native solutions are gaining vertical momentum. Vertical incumbents are no less at risk. In fact, the stakes are higher - industry-specific Jobs to be Done tend to be even more customized and critical, and they are no longer beyond the reach of a specialized model.

Ultimately the companies that survive this disruption won’t be the ones that hold their customers hostage with high switching costs. They will be the ones that leverage AI to stop being passive “Systems of Record” and start being active “Systems of Judgment.”

In the end, it won’t be the stickiest who survive—it will be those who are most useful.

Leave a comment