Anthropic prices its IPO next week on the assumption that AI spending keeps compounding. Today that spending comes from a narrow band of technical companies, and it has to reach the rest of the economy before the money financing the build-out runs short. That is the problem LazyFox exists to solve.
The build-out only pays for itself if spending on AI keeps compounding, and the demand side is the part nobody can see clearly.
Anthropic is expected to price its IPO on November 18. Anyone buying that stock is also buying into the infrastructure around it, and that infrastructure is enormous. Alphabet, Amazon, Meta, Microsoft and Oracle are expected to spend roughly $800 billion this year and more than $1 trillion in 2027 on data centers, chips and power.
The supply side of that bet is easy to observe. Chipmakers, cloud providers and data center builders are mostly public companies that report every quarter, and almost every AI chip installed today is rented out at a premium. The demand side is much harder to read. Anthropic's amended S-1 has given the market its first audited look at a model company, with almost $17 billion of revenue in the third quarter alone, up more than elevenfold on a year earlier. But OpenAI remains private, and no official statistic measures what businesses across the economy spend on AI models.
That makes card and expense data some of the best evidence there is. Ramp sees the spend of more than 85,000 US businesses, including more than $10 billion a year of AI spend, and its lead economist, Ara Kharazian, has spent the last three months publishing a series he calls "cracks in the AI thesis." Read next to the financing side of the build-out, his numbers describe a market with real demand, a narrow buyer base and a timing problem.
Every data center financed today is a bet that companies outside tech will pay for AI within a few years. Most of them can't use it well yet.
LazyFox exists because of that second sentence. We don't think the demand is fake, and we don't think the build-out is a railway to nowhere. We think the next wave of demand depends on a missing piece of enterprise infrastructure, and that it has to arrive faster than the market currently assumes.
AI demand runs deep in a few technical sectors and thin almost everywhere else.
Ramp's latest index puts paid AI adoption at 56.8% of US businesses in October, 13.3 points higher than a year earlier. Adoption is broad, but the money behind it is not: about 80% of OpenAI and Anthropic spend comes from 1% of firms.
Kharazian's breakdown per employee shows how steep that curve is. Among businesses that pay for AI, the median company spends $12 to $15 per employee per month. A company in the top 10% spends around $800, and one in the top 1% spends upwards of $10,000.
Concentration on its own would be ordinary. The top percentile of advertisers drives most digital ad spend too. What Kharazian flags as unusual is how alike the top 1% are. They are highly technical companies in software and media, finance and technical services, so their budgets tend to rise and fall together, and a correction in those sectors would hit the model companies' revenue all at once.
Coding is what built that concentration. Coding agents were the first AI product companies would pay thousands of dollars per engineer for, because the return shows up in shipped software every week, and the revenue that made Anthropic an IPO candidate came largely from that one use case.
Usage keeps rising at the top 1%. What they pay for each unit of it keeps falling.
Top-1% spend per employee has barely moved since the summer: about $9,900 in July and about $10,200 in October, after a year in which it had grown much faster. Ramp is clear that this is not a loss of faith in AI. Those firms use more AI every month and buy it from more vendors. They have also learned to spend less on it.
Three forces drive that. Frontier prices keep falling as the labs compete for share, with extra discounts for long contracts, and the effective price per million tokens fell 41% between March and August. Companies have also taken control of their AI bills: they route simple tasks to cheap models, reserve frontier models for hard ones and block expensive models where they add nothing. And because AI spend is spread across providers instead of locked into a single contract, a buyer can move volume to whoever is cheapest next month. Anthropic's own filing shows the effect: price cuts on new models and contract discounts took about $600 million off its third-quarter revenue, and sequential growth slowed from 144% in the second quarter to 47% in the third.
Model quality has limits as a pricing lever too. Kharazian points to the most capable model released this year, which topped out at 5 to 10% of business spend. A close rival was nearly as good and a little cheaper, and it came without the data-collection requirements most enterprises could not accept, so businesses chose the model that fit their obligations and their budget.
The companies already getting value from AI are getting better at buying it. That helps them, and it squeezes the labs.
For the labs, the squeeze comes from both directions at once. Their most engaged customers are optimizing, and most of the rest of the economy has barely started spending.
First free cash flow, then bonds, then equity, then pension and insurance money. Each step puts more pressure on demand to show up soon.
The first stage of the build-out was paid for out of free cash flow, and within about a year that stopped being enough. The five largest US tech companies issued an average of roughly $28 billion of bonds a year between 2020 and 2024. In 2025 they issued $121 billion, and by mid-2026 they had already issued $159 billion, against a Bank of America forecast of $175 billion for the full year.
Equity came next. In June, Alphabet raised $84.75 billion in the largest equity transaction in corporate history and its first equity raise in more than two decades, after years as one of the biggest buyers of its own stock. In August, Nvidia signed agreements with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to build financing platforms meant to mobilize more than $500 billion for its customers' data centers, much of it from pension funds, insurers and sovereign wealth funds.
Each rung down that ladder is further from the operating business that ultimately has to repay the money. There is also a lag built into the whole system. Chips, power and buildings take years to bring online, so the capital raised in 2026 is a bet on demand near the end of the decade.
None of this means AI is a fad. We think its economic impact will be very large. The risk is a gap in time. The financing has to keep rolling forward until AI revenue can fund the build-out from cash flow again, and if demand arrives later than the financing assumes, that gap is where the bet breaks.
You can believe AI will change the economy and still worry that the bridge to its returns is too short.
The only pool of money large enough is labour and productivity, which is where the evidence is thinnest.
Spending on AI models runs at roughly $200 billion this year. For the bet to work, it has to reach about $400 billion next year and around $800 billion the year after. Hyperscaler infrastructure spending alone already exceeds $800 billion a year.
Existing software budgets can't close that gap. Companies already pay for licences, and moving some of that money to AI tools reshuffles spend without creating much that is new. Amounts on this scale can only come from work companies currently pay people to do: labour that AI takes over, or output that rises enough to justify thousands of dollars per employee.
That is where the evidence is thinnest. In an NBER survey of executives across the US, UK, Germany and Australia, nearly 90% of firms reported no AI impact on productivity or employment over the previous three years, while expecting a 1.4% productivity gain over the next three. Kharazian expects AI's effect on the economy to be meaningful but modest, closer to what computers delivered: a gain of one or two percent a year, real and still small next to the valuations built on top of it.
So the build-out depends on productivity gains showing up quickly, and in companies that look nothing like the early adopters. If large traditional enterprises can't show a return, they will stop paying, or never start paying at the scale the infrastructure assumes.
Consumers, the office suite and engineers placed inside customers are all attempts to widen the buyer base.
The model companies see the same numbers, and their response so far points in three directions at once.
The third direction tells you the most. A forward deployed engineer sits inside a customer for months, learning how the work gets done, which systems hold which numbers and what the business means by its own terms, and then builds the AI around that knowledge. Nobody pays for that when a product is easy to adopt. The billions going into FDEs are an admission that model capability has run far ahead of what the average company can absorb.
Forward deployed engineers are the labs' answer to a context problem, and people don't scale like software.
FDEs work, as Palantir showed over a decade. They are also expensive, slow to staff and limited in number. Ten thousand engineers spread across the companies that make up the rest of the economy would barely reach most of them, and the knowledge each engineer builds up tends to stay inside that engagement.
The gains are real at companies that use AI intensively. Light users look like non-users.
The productivity gains are real. Ramp found that firms adopting AI grew headcount by about 10% in the two years after adoption, so they were hiring more people, not fewer. The effect was concentrated entirely in businesses that used AI intensively: at least $50 per employee per month, coding agents, several models, sustained over time. Businesses that used it lightly, which describes the average AI spender, showed no difference from the control group.
That points at where the real dividing line sits. It isn't access to models, which every company now has, and it isn't price, which keeps falling. It is whether AI can be trusted with the work that matters inside that particular company, and the firms above the line are the ones where it can.
Kharazian hears the same thing from businesses that hold back. Large traditional enterprises have the budgets, but they won't roll an AI system out to thousands of employees or millions of customers until they are sure it works reliably at scale. Once that hurdle is cleared, the next question is how to control the cost. Both are questions about trust before they are questions about models.
A coding agent works inside a system that already defines meaning and checks results. A finance agent has nothing like it.
The usual explanation for coding becoming the first breakout use case is that engineers adopt new tools early. That's true, but it misses the structural reason. Code carries its own context. A repository defines what every function and variable means, a compiler rejects what doesn't fit, tests show whether a change did what it claimed, and a reviewer can read the diff and approve it. An agent working there finds out it is wrong within seconds.
Business work has none of that. Ask an agent for last quarter's revenue and it finds bookings in Salesforce, invoiced revenue in SAP and recognized revenue in the warehouse. Which one counts depends on who is asking and why, and that knowledge sits in people's heads, in spreadsheet tabs and with a few long-serving employees. The output looks plausible, the controller doesn't trust it, and usage stalls.
That is why coding agents pay off for the roughly 3% of Americans who write software, while the gains haven't followed into the much larger share of the workforce whose work runs on business data.
Not every job will use AI heavily. A typical restaurant or doctor's office is not going to spend thousands of dollars per employee on it. Between software engineers and cashiers, though, sit tens of millions of people in finance, operations, sales, procurement and planning at large companies. That group is big enough to carry the next wave of demand, and its work runs on exactly the kind of data agents can't interpret today.
Give every agent in the enterprise the governed understanding of the business that a repository gives a coding agent.
LazyFox's mission follows from all of this. We want AI to deliver on its promise in the companies that make up most of the economy, not only in the technical firms that already run on it, and we want that to happen within the window the build-out allows rather than a decade later.
That means giving finance, operations and sales agents what coding agents get from a repository: one place where the meaning of the business lives, written down, governed and checkable. We call it an Enterprise Intelligence Layer, and it has to solve three problems at the same time.
LazyFox connects read-only to the systems a company already runs and reconciles meaning across all of them, without migrating data. It works on three levels.
Because meaning is resolved before a question reaches a model, 60 to 80% of routine enterprise queries run directly from governed code and never consume model tokens. Answers arrive with the definition and the query behind them, under row- and column-level permissions, so a controller can check them the way a reviewer checks a diff. And because the layer is model independent, a company can move between Claude, GPT and open-weight models without rebuilding its institutional knowledge.
Customers see the effect in work AI rarely reached before. Finway, a Munich-based spend management platform, builds new reports in hours instead of weeks and works from one definition instead of two conflicting legacy reports. Another customer replaced Power BI with it entirely.
The labs are paying engineers to carry context into each customer by hand. We think that context belongs in software, built once and owned by the company.
The model companies will keep making models more capable. The enterprises that ultimately have to pay for the build-out mostly can't use the capability that already exists, because their agents don't know what their data means and their people can't verify the answers. Close that gap and a large traditional enterprise starts to use AI the way the top 1% does today. That is the work LazyFox exists to do, and the timing is what makes it urgent.
Falling prices are on your side. What holds you back is whether you can trust what AI produces.
Companies that do this over the next two years will get AI at falling prices, with returns they can show their boards. Demand from companies like them is also what the build-out needs to pay for itself.
The AI infrastructure build-out (roughly $800 billion of hyperscaler capex in 2026 and more than $1 trillion expected in 2027) is financed on the assumption that spending on AI models keeps compounding, from about $200 billion in 2026 to $400 billion in 2027 and $800 billion in 2028. Ramp data shows that about 80% of OpenAI and Anthropic spend comes from 1% of companies, concentrated in software, finance and technical services; median businesses spend $12 to $15 per employee per month, the top 1% around $10,000; top buyers are optimizing as the price per million tokens fell 41% between March and August 2026. Financing has moved from free cash flow to record bond issuance, Alphabet's $84.75 billion equity raise and Nvidia's $500 billion financing platforms drawing on pension and insurance capital.
The post argues that coding agents succeeded because code carries its own context (repository, compiler, tests, review), while business agents in finance, operations and sales lack governed meaning, so their output cannot be trusted. It sets out LazyFox's mission: provide an Enterprise Intelligence Layer covering data, meaning and trust, read-only and without data migration, so AI delivers productivity gains beyond the early adopters before the gap between infrastructure financing and AI revenue becomes the failure point. It closes with five steps for enterprises outside tech.
LazyFox gives every agent the same governed meaning of your business, read-only on your existing stack and without data migration.