Amazon passes Walmart, and the engine behind it
The hosts opened with the news that Amazon has overtaken Walmart as the largest company in the world by revenue. Amazon reported $716.9 billion for its most recent fiscal year, narrowly ahead of Walmart’s record $713.2 billion, which ended Walmart’s roughly 13-year run at the top of the Fortune 500.
Andy focused on what powered the result. A meaningful and growing share of Amazon’s revenue comes from businesses that are not first-party retail, chiefly AWS and a fast-growing advertising arm, where Walmart’s revenue is still overwhelmingly retail. He tied this to a habit of mind: Amazon has long built infrastructure for itself and then sold it to others, including running Target’s website from 2001 to 2011 and turning its internal compute into AWS. Walmart, by contrast, did not launch Walmart Fulfillment Services until 2020. Brian’s read was that older, larger companies tend to protect their existing business and move slowly, while Amazon has been willing to compete with itself, as it did when it opened its marketplace to third-party sellers two decades ago. The throughline for both hosts was to follow the money and model the opportunity rather than dismiss it.
AI sticker shock: the bill nobody was watching
The main topic grew out of a pattern in recent headlines. Companies have been pushing employees hard to adopt AI, in some cases mandating it, building internal usage competitions, and tying it to performance reviews. Reporting from the Wall Street Journal describes firms including Meta, Microsoft, and Salesforce pressing employees to use AI, and the term “tokenmaxxing” has emerged for the race to consume as many tokens as possible. The catch is that few finance teams were tracking the resulting spend.
Three years ago, would you or I have thought we would be doing a podcast titled, are AI token costs getting out of control?
Andy Hoar, Master B2B
The example that set off the conversation came from an Axios report in late May, which said AI sticker shock had hit corporate America. It described a single consultant’s client that reportedly spent about $500 million in one month on Claude after failing to put usage limits on employee licenses. Both hosts noted that the company was never named, and the figure has since been questioned, with at least one analysis arguing that a bill that size would require deliberate, large-scale use rather than a forgotten setting. The more verifiable cases point the same direction. Uber reportedly burned through its 2026 AI budget within a few months, with its leadership saying the spend was getting hard to justify, and Microsoft has scaled back internal Claude Code licenses over cost.
How big the token numbers have gotten
To scope the scale, Andy cited Google’s own disclosure that it processes about 3.2 quadrillion tokens a month, roughly 40 quadrillion a year and about a sevenfold increase from the year before. He offered a way to picture it: counting one number per second, reaching 40 quadrillion would take well over a billion years. The driver behind the spend is agentic AI, which repeats queries in sequence and can consume far more tokens than a single chatbot question. Goldman Sachs has projected that token consumption could grow many times over by the end of the decade.
Why token usage is a poor measure of value
The hosts kept returning to the ROI question, and their answer was that token usage is a weak proxy for it. If employees are rewarded for consumption, they will use AI for everything, including trivial tasks, which inflates the bill without adding value. Andy described a four-person startup that reportedly spent about $113,000 a month on tokens and drew criticism for it, while the founders argued it was working for them. His point was that the return depends on the use case, the speed gained, and the revenue produced, none of which a raw token count captures. Reporting since has reached a similar conclusion, that the value of an output has little to do with how many tokens it took to generate.
A familiar pattern: infrastructure buildouts and falling prices
Andy framed the current moment against past infrastructure cycles. He pointed to the late-1990s fiber-optic buildout, when companies overbuilt capacity and the price of a T1 line fell sharply, contributing to bankruptcies such as Global Crossing. He expects token prices to follow a similar path downward as supply rises to meet demand. Brian connected the spending by Amazon, Google, and others, including Amazon’s plan to spend up to $200 billion on AI infrastructure in 2026, to earlier bets on data and communications infrastructure. The shared view was that prices come down once the capacity is in place.
What it means for your ecommerce platform bill
For practitioners, the practical question Brian raised is what happens to the cost of the platforms they already run, since commerce, search, and marketing tools are all embedding AI. Andy’s prediction was a shift in how these tools are priced.
You can’t have a variable model and a fixed model. It has to be a consumption-based model.
Andy Hoar, Master B2B
He expects pricing to move from fixed subscriptions toward consumption, then moderate and stabilize, likely with caps and tiers once both sides get more disciplined. He used an energy analogy to explain the current overuse: if a new power source were offered for a flat twenty dollars a month with unlimited use, people would leave every light on. Flat, unlimited AI plans create the same incentive, which is why he expects metered, consumption-based pricing to win out. The takeaway for B2B ecommerce leaders is that AI cost is becoming a P&L line worth forecasting and governing, rather than a fixed software fee to set and forget.
The real open question is usage
The hosts closed on the tension that will define the next few years. The cost per token is widely expected to fall, but usage is expected to climb as teams find more uses for AI.
The water will find the equilibrium, but the price is going to come down. The question is how the usage is going to go.
Andy Hoar, Master B2B
If per-token prices drop while consumption keeps rising, total spend may not fall much even as the unit economics improve. For now, the discipline both hosts recommended is straightforward: track where AI is creating value, set sensible limits, and avoid using expensive tools for tasks that do not need them.