Amazon's Anti-Tokenmaxxing Memo Puts $220 Billion of AI Capex on Trial, Two Days Before Earnings

On October 27, 2026, Amazon published an updated role guideline for software development engineers that tells managers to stop measuring AI adoption by token counts. The memo calls token usage, the rate of AI-tool adoption, and code changes or reviews \"deeply flawed\" ways of measuring teams, warns that \"tokenmaxxing is at odds with frugality,\" and instructs leaders to judge work by long-term outcomes instead. Amazon has replaced the token count with the shipped outcome as its internal scorecard — a reversal for the company that scrapped an internal leaderboard ranking engineers by tokens burned only five months earlier.

The timing is the signal. Amazon reports third-quarter results on October 29, and the guideline landed two days earlier, in a quarter where the spending curve has not bent. Capital expenditure ran $54.2 billion in the second quarter against $32.1 billion a year earlier, and trailing free cash flow has flipped to an outflow of $7.6 billion from an inflow of $18.2 billion. Guidance for the third quarter calls for net sales of $197 billion to $202 billion, or 9% to 12% growth, against 20% growth in the second quarter. Amazon is asking its own engineers to squeeze more out of a capital base it is still expanding.

Three links connect an engineering guideline to an income statement. First, wasted agent loops still consume AWS accelerator hours — hours that paid customer workloads could have used — and CEO Andy Jassy said on the July 30 call that capacity demand through 2028 will not be met even at $220 billion of spending. Second, Amazon is paid for compute, storage and network, not for tokens; a customer that burns fewer tokens but runs more tasks still consumes the same silicon. Third, with capital spending running ahead of operating cash flow, utilization rather than installed capacity determines whether the build earns its cost of capital. Cutting internal token waste protects AWS margin and cash flow without losing a single customer.

Here is what the memo exposes about who keeps the money. Micron turned $133.19 billion of fiscal 2026 revenue at an 80.7% GAAP gross margin, with its core data center business unit at $18.0 billion of revenue and a 90% gross margin in the September quarter. SK hynix posted 79.3 trillion won of second-quarter revenue, up 257% year over year, and 60.5 trillion won of operating profit — a 76% operating margin. NVIDIA grew data center revenue 117% year over year to $89.0 billion in its June quarter. Amazon, meanwhile, ran $200.6 billion of quarterly revenue at a 13.7% operating margin and told investors the single biggest reason to spend $20 billion more was the price of memory. The margin in this stack accrues to the layer selling bits of silicon, not the layer selling tokens.

The metric shift is a pricing event, and it lands hardest on the labs. If the unit of value becomes a completed task rather than a token, metered revenue per task falls unless the price per token holds — and price per token has been falling all year. Anthropic grew revenue 45% sequentially to $16.8 billion in the September quarter while profitable on an adjusted basis. Volume was never the problem; the pricing unit is — labs that keep selling tokens while buyers move to outcome budgets will see revenue per task compress even as token counts climb.

Amazon does not earn more when a customer burns more tokens; it earns when a customer runs more workloads on AWS capacity. In the quarter Amazon added six frontier and open-weight model families to Bedrock and shipped TorchNeuron, which lets existing PyTorch code run on Trainium without a rewrite, it widened the number of ways to put a task on its own silicon. Amazon wants more tasks per unit of HBM, not more tokens per task — and the memo tells engineers to optimize for exactly that.

The next test arrives almost immediately. Amazon reports third-quarter results on October 29, 2026; SK hynix reports its third quarter on October 28. Two consecutive reports will show whether token discipline lands before or after the cost of the memory that powers it — because for Amazon, both sit on the same balance sheet.

Over one to three years, the memo is a leading indicator of how AI gets billed. If outcome-based pricing spreads — per task, per resolution, per shipped change — metered tokens become an input cost line for buyers and a pass-through for labs, and the winner is whoever delivers the cheapest completed task per dollar of HBM. Amazon has positioned itself on both sides of that trade: Trainium for cheap inference, Bedrock for distribution, and a stake in the model layer through Anthropic. Longer term, the question is not how many tokens get sold but how many tasks get completed per unit of compute.