AI companies can improve margins without throttling product growth by lowering the cost of delivering each useful result, aligning prices with usage or customer value, and making delivery more repeatable. The key is to track those gains alongside quality, latency, adoption, retention, and customer outcomes: cost cuts that weaken the product can undermine the growth they are meant to support.
Start with the economics of a useful result
AI infrastructure costs are not a single fixed overhead number. They vary by model, feature, customer, and workload. Start by measuring cost at the level where the business creates value: a query, completed task, feature, or customer. Where possible, distinguish successful work from calls that fail, are retried, or produce little customer value.
As an Amazon Associate I earn from qualifying purchases.
Then compare the cost of serving that work with the revenue it supports. A lower cost per query is useful only if the result still meets the product’s requirements. Track unit cost next to task quality, latency, reliability, and customer usage so optimization does not quietly turn into a worse experience.
Free tools Windows power users keep installed
One-click scans. No signup required.
Test model and infrastructure choices by workload
For each workload, assess model choice and routing, architecture, batching or scheduling, and resource utilization. A smaller or cheaper model may work for routine requests while a more capable model remains necessary for harder tasks. Routing can help, but only if it sends work to a model that meets the feature’s quality and reliability bar.
#1 Best Overall
Compare options using total cost for the workload, quality on the intended task, latency, reliability, utilization, and deployment constraints. There is no standardized apples-to-apples vendor benchmark in the cited evidence, so test with representative workloads rather than assuming a published comparison predicts your own economics.
In its 2026 survey of software-company builders, ICONIQ reported that two-thirds of respondents had improved per-query unit economics. Respondents cited inference-cost management, model routing, and revenue growth creating cost leverage; this is reported experience, not controlled evidence that any one lever caused the improvement.
A company filing with the Hong Kong Exchanges and Clearing Limited offers a specific example, not a sector benchmark: inference cloud-service costs accounted for more than 90.0% of that issuer’s cost of sales in each year of its track record period. The issuer described model-architecture improvements, dynamic resource allocation, a unified training-inference framework, and higher utilization as ways it worked on efficiency. Its cost of sales was 124.7% of revenue in 2023, 87.8% in 2024, and 76.7% in the nine months ended September 30, 2025; its AI-native product gross margin moved from negative 23.5% to 4.7% between the nine months ended September 30, 2024 and September 30, 2025. These figures describe one issuer and different reporting periods, not a typical AI-company trajectory. (HKEX filing, 2025)
Choose pricing that fits usage and value
A fixed subscription can make purchasing and revenue planning straightforward, but it may poorly match costs when customer usage varies widely. Consumption pricing can connect the bill more closely to workload; outcome-based pricing can tie it to value delivered, but requires a credible way to define and measure that outcome. A hybrid can preserve a predictable entry price while charging for heavier usage or higher-value work.
Pricing is moving toward a mix rather than one universal model. ICONIQ’s 2026 survey reported that consumption-based pricing increased from 35% to 42% over six months and outcome-based pricing from 18% to 23%; surveyed companies blended an average of 1.7 pricing models. Those are survey findings, not proof that a particular mix will improve margin or suit every customer.
Before changing packaging, model how the design affects adoption, usage incentives, gross margin, and revenue predictability. In particular, check whether customers can get value without hitting a usage limit that discourages adoption, and whether the price for high-volume or high-complexity use covers its cost to serve.
Rank #3
Make delivery repeatable before adding more customization
Revenue scales more efficiently when each new customer does not require a fresh implementation. Standardized product offerings, reusable software and hardware components, and repeatable deployment configurations can reduce engineering effort and delivery complexity. Keep customization for cases where customer value and strategic learning justify the extra implementation and support work.
A 2026 HKEX filing describes one issuer’s strategy of prioritizing higher-value engagements while expanding standardized offerings, with reuse intended to reduce delivery effort and complexity. That is the company’s stated rationale, not causal proof that standardization alone produced its results. (HKEX filing, March 16, 2026)
To decide which work to standardize, compare customer value and contribution margin with implementation time, ongoing support effort, and repeatability. Reusable components can improve economics; refusing all customization can also mean turning away valuable customers or learning.
Separate gross-margin improvement from operating leverage
Gross margin and operating margin answer different questions. Gross margin reflects the economics of delivering products and services; operating margin also includes expenses such as research and development, sales and marketing, and general and administrative costs. A company can make progress on the latter as revenue grows even while AI infrastructure and product usage continue to weigh on the former.
For context, ICONIQ’s 2026 survey of software-company builders reported gross margins of 45% in 2025 and projected 53% in 2026 and 59% in 2027. Its survey also put AI products at 32% of revenue in 2025, projecting 42% in 2026 and roughly 53% by 2027. The 2026 and 2027 figures are projections, and none is an audited industry total.
Reported company results show why those measures should not be conflated. Microsoft reported a Microsoft Cloud gross margin percentage of 66% in FY2026 Q3, saying AI infrastructure investment and growing AI product usage were downward pressures, partly offset by efficiency gains in Azure and Microsoft 365 Commercial cloud. Microsoft company-wide operating income grew 20% year over year in the same quarter. The scopes differ: the Cloud figure is not Microsoft’s company-wide gross margin. (Microsoft FY2026 Q3 results)
Best Value
Infrastructure investment can also bring costs beyond immediate usage. Alphabet said nearly 75% of Google Cloud customers had used its vertically optimized AI offering, and those AI customers used 1.8 times as many products as customers who had not used AI. In the same 2025 Q4 earnings call, it said investment in infrastructure was increasing depreciation and data-center operating costs such as energy; depreciation rose by nearly $6 billion, or 38%, from $15.3 billion in 2024 to $21.1 billion in 2025. These are Alphabet’s company-reported figures, not a general estimate of AI customer value or infrastructure cost. (Alphabet 2025 Q4 earnings-call transcript)
Hardware strategy is another long-horizon decision. Amazon said Trainium3 was 30–40% more price-performant than Trainium2 and expected several hundred basis points of AWS operating-margin advantage at scale compared with relying on others’ chips for inference. These are management’s claims and expectations, not independently verified realized savings; the expected advantage depends on reaching scale. (Amazon 2025 letter to shareholders)
Track spending ratios without cutting indiscriminately
Monitor research and development, sales and marketing, and general and administrative costs against revenue, while identifying the investments that support adoption and differentiation. A 2026 HKEX filing reported one issuer’s overall gross profit margin at 30.5% in 2023, 32.3% in 2024, and 37.3% in 2025. It also reported adjusted total operating expenses, excluding share-based payment expenses, at 113.6%, 83.4%, and 63.9% of revenue in those years. The distinction matters: reported 2025 total operating expenses were 107.7% of revenue, with share-based payment expenses a material factor. These are issuer-specific figures; the adjusted ratio is not the same as reported total operating expenses. (HKEX filing, March 16, 2026)
Outdated Drivers Are Slowing You Down
One free scan finds every outdated or missing driver and matches the right update for your exact hardware.Free scan · exact hardware matchPC Slower Than It Used to Be?
A free scan shows the junk files, broken settings and background clutter dragging Windows down - then fixes them in one click.Free scan · Windows 10 & 11Use a decision process that protects the growth engine
- Define the unit of value. Pick a query, successful task, feature, or customer segment that maps to customer benefit and can be measured consistently.
- Establish the full cost to serve. Include relevant inference and infrastructure usage, deployment, and support effort; split costs by workload or customer where the data permits.
- Test one lever against a baseline. Evaluate model routing, architecture, utilization, packaging, or standardization on representative work rather than assuming it will save money.
- Set guardrails before rollout. Track quality, latency, reliability, adoption, retention, and customer outcomes alongside unit cost and margin.
- Scale only when the trade-off holds. Check that savings or higher realized revenue persist at greater usage and across the customers the change is meant to serve.
These steps help distinguish genuine efficiency from cost displacement—for example, a cheaper inference path that creates more retries or support work—or from a short-term margin improvement that reduces product value. The cited surveys and company disclosures do not prove that any particular cost action is harmless to growth.
What the evidence can—and cannot—tell you
The available examples span survey respondents, individual company filings, and earnings disclosures with different scopes, definitions, and periods. Their percentages should not be compared as if they came from a common measurement framework. They show that inference economics, pricing, repeatability, and operating leverage are active management questions; they do not establish a universal margin target, a best pricing model, or causal proof that the described levers preserve product growth.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
