AWS's October to December 2026 operating margin, alongside Amazon's cash capital expenditure, is the next test of operating leverage.
An agent on a schedule has acquired a peculiar privilege: it can spend a computation budget without asking again. LangChain's 17 November 2026 account says more than 70% of runs by its own agents built on Managed Deep Agents now begin on schedules rather than ad hoc human prompts. A standing schedule shifts the spending decision from a person with a question to an instruction that repeats. Tomorrow's run arrives whether today's output changed anything. The supplier pays to serve the request before anyone decides whether the work earned its keep.
We would back operators turning scheduled demand into paid utilization, rather than the company with the busiest run counter. At AWS, our specific wager is an operating margin above its Q3 2025 level while Amazon's capital spending rises. Who Keeps the Margin is our rule: the saving from cheaper inference belongs to the party controlling the workload, the customer contract or the serving default. Its test is operating income as a share of sales against the capital bill. A cheaper request helps its buyer; profitable repeat business rewards the infrastructure owner.
LangChain's 13 November 2026 funding announcement put $300M of new capital beside a $3B valuation and said more than 7,000 organizations use LangSmith. The valuation prices an opportunity, while LangSmith adoption and scheduled Managed Deep Agents runs measure different things. A standing assignment promises recurring demand without another sales pitch for each task. It also gives a mediocre assignment the power to repeat itself. An agent that avoids a worthless run saves the whole request; making a necessary run cheaper saves only part of its cost.
Amazon supplies the other half of the story. Its 16 November 2026 SageMaker AI announcement introduced KV cache-aware routing for real-time GPU inference endpoints using the AWS vLLM container, at no additional charge. The feature needs at least two instances or inference-component copies, so the router has a destination to choose. With a repeated prompt, it favors an instance that already holds the relevant cached computation. Think of a factory sending a repeat order back to the machine still set up for the job. LangChain's agents and SageMaker's router are separate products; repetition connects their economics.
In its 17 November 2026 benchmark, AWS compared the new routing with random routing on Llama 3.1 70B Instruct across seven ml.p5.48xlarge instances. AWS measured up to 97% less median time to first token and up to 234% higher throughput; cache hit rates went from 3% to 48% under random routing to 80% to 93% with cache awareness. Those are AWS's own best results for selected long-context workloads. For AWS, the available capacity earns more only when it serves additional paying demand.
At unchanged traffic, a 234% throughput gain makes a fleet scenario of one divided by 3.34, or 30% as many instances. The 10 September 2026 prefix-aware routing result of up to 16% higher throughput produces one divided by 1.16, or 86% of the fleet under the same scenario. These workloads differ, and fleet ratios are scenarios rather than cash savings. A customer choosing fewer instances captures the saving directly. AWS instead captures value by selling more requests on capacity it already operates, provided the extra revenue turns into operating profit. Who Keeps the Margin depends on that allocation, with contract prices setting how much customers retain.
Epoch AI's 12 March 2025 analysis supplies the toughest case against making routing the main economic story. It found that the price of matching GPT-4 on PhD-level science questions "fell by 40x per year". Its MMLU comparison runs from GPT-4-0314, scoring 86.0 at $37.50 per million tokens in March 2023, to Gemini 2.0 Flash, scoring 88.0 at $0.18 in February 2025. The $37.50 price divided by $0.18 is a 208-fold decline at similar measured performance. A cheaper model serves fresh prompts as well as repeated ones, while caching needs context worth reusing. On the unit-price question over that period, models win.
A routing gain can also evaporate on the wrong job. Huang Cheng's 20 August 2026 CacheRoute study found two 32B workloads with a failure mode: "when affinity recovers too little KV work, its residual load skew reduces or erases the improvement." On those jobs, sending requests toward a cache can sacrifice balanced traffic for too little recovered work. Yet the same study measured 2.3 times the strongest baseline's serving capacity on a 70B workload across 60 H100 GPUs. The two results come from the same research, so the sensible response is to replay a customer's actual traffic before assigning value to cached work. A useful router earns its place in profitable serving, workload by workload.
Jevons-style rebound is the uncomfortable historical pattern here: making each unit cheaper can leave the total bill larger. Epoch's 208-fold reduction in model price between March 2023 and February 2025 sits beside Amazon cash capital expenditures increasing from $34.2 billion in Q3 2025 to $75.9 billion in Q3 2026, a rise of $75.9 billion over $34.2 billion less one, or 122%. Cheaper units and a growing total bill are the resemblance. The periods differ, the capex is Amazon-wide rather than solely for inference, and the price change does not establish a causal link; AWS margin also rose. The dangerous outcome is a capital bill growing faster than profitable demand.
An expanding capital budget can coexist with a richer operating margin. Amazon's filing for the quarter ended 30 September 2026 puts AWS sales at $33,006 million in Q3 2025 and $45,984 million in Q3 2026, while operating income rose from $11,434 million to $17,773 million. Income divided by sales went from 11,434 over 33,006, or 34.64%, to 17,773 over 45,984, or 38.65%. This increase preceded November's routing launch, making it a baseline for operating leverage rather than a feature return. The larger sales base carried a larger profit share, even before the routing feature existed.
The contracts make that distinction urgent. At 30 September 2026 Amazon held approximately $578 billion of contracted future services, primarily related to AWS, with a weighted-average remaining life of 6.1 years. The company expects cash capital expenditure to increase meaningfully in 2027. Amazon's filing says AWS's 39% Q3 2026 sales growth came mainly from greater customer usage, partly offset by pricing changes driven primarily by long-term contracts. The contracted work creates a reason to build; price terms determine how much of each efficiency gain the buyer keeps. A scheduled agent is valuable to AWS only when the revenue from useful recurring work exceeds the expense of serving it.
We would back operators only while paid recurring work preserves operating margins through rising capital bills, and pass on businesses priced chiefly on scheduled runs. Amazon's fiscal 2026 Form 10-K, due 1 March 2027, would prove us wrong if AWS's October to December 2026 operating margin falls below 34.64% while Amazon cash capital expenditure keeps rising, a test of aggregate operating leverage rather than the routing feature alone.