AI companies can improve margins without choking off product growth by lowering the cost of delivering useful results, matching pricing to customer value and usage, and making delivery more repeatable. The key is to measure unit economics alongside quality, latency, adoption, and retention: a cheaper product that customers use less is not a durable margin improvement.
Start by locating where the economics are changing
AI growth can expand revenue and costs at the same time. Inference, cloud capacity, depreciation, and energy can rise with product usage, while research and development, sales, and support costs determine how much revenue converts into operating income. Gross margin and operating margin therefore answer different questions: the first captures direct delivery costs; the second also reflects operating expenses.
As an Amazon Associate I earn from qualifying purchases.
Track costs at a useful level of detail—by query, feature, customer, workload, or successful task—rather than treating infrastructure as one undifferentiated bill. Pair each cost measure with its product counterpart: task quality, latency, reliability, usage, renewal, or customer outcome. That makes it possible to spot a cost reduction that is actually shifting expense to support, hurting reliability, or reducing adoption.
Quick wins for a faster PC:
Scan for outdated or missing drivers - takes under a minuteDriver Scan →Repair Windows errors before they cause bigger problemsFix Now →The available company examples illustrate why margin paths vary. A 2025 Hong Kong Exchanges and Clearing filing reported that inference cloud-service costs exceeded 90.0% of cost of sales in each year of the issuer’s track-record period. 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; for its AI-native product, gross margin moved from negative 23.5% in the nine months ended September 30, 2024 to 4.7% in the same period of 2025. These are figures for that issuer and those periods, not sector benchmarks. Read the filing.
#1 Best Overall
Lower inference cost while preserving the result customers need
Inference optimization is a product and infrastructure decision, not simply a mandate to use a smaller model. Test changes against representative workloads and customer requirements, then compare the full cost of serving the task with the quality and speed it delivers.
- Route by task. Use a more capable model where complexity or risk warrants it and test less costly options for routine requests. Measure whether routing changes accuracy, fallback rates, latency, and cost per successful task.
- Improve architecture and utilization. Evaluate batching, scheduling, dynamic resource allocation, and other architecture changes against peak and typical demand. A cheaper configuration that creates queues or unreliable service may undermine adoption.
- Match capacity to demand. Review utilization and deployment constraints by workload. Better allocation can reduce idle spend, but test it against response-time and availability requirements.
- Recheck the economics as usage changes. A product feature that looks inexpensive at launch can become a material cost center as customers use it more or as workload mix shifts.
In ICONIQ’s 2026 survey of software companies building AI products, two-thirds of surveyed builders reported improved per-query unit economics. Respondents cited inference-cost management, model routing, and revenue growth creating cost leverage as contributors. This is reported experience, not a controlled demonstration that any one method causes savings or protects growth. See ICONIQ’s State of AI: The Builder’s Economy.
Price for usage and value, not just access
Pricing design affects both customer adoption and cost recovery. A subscription can make spend predictable, while consumption pricing can better reflect variable usage; outcome-based pricing can tie payment more closely to delivered value. A hybrid can preserve an accessible entry point while charging more as usage or customer value rises. None is automatically right: the best design depends on willingness to pay, variability of usage, cost-to-serve, adoption friction, and revenue predictability.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Before changing packaging, model how the change affects who tries the product, how much they use it, what it costs to serve those customers, and how readily they can expand. Make sure customers can understand and forecast the bill. Monitor whether usage pricing discourages beneficial use or whether a flat plan invites workloads that exceed the price’s economics.
ICONIQ reported that consumption-based pricing rose 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. The figures describe survey respondents, not all AI businesses, and do not establish a universally superior pricing mix. ICONIQ’s 2026 report also said AI products accounted for 32% of revenue in 2025, with projections of 42% in 2026 and roughly 53% in 2027. The 2026 and 2027 figures are projections, not realized industry totals.
Make delivery repeatable as revenue scales
Revenue growth improves margins only when the work required to win, implement, and support each customer does not grow at the same rate. Standardize the parts of delivery that recur: product configurations, integrations, deployment patterns, documentation, and support workflows. Keep bespoke work for cases where its customer value and strategic learning justify the engineering and support burden.
Rank #3
A 2026 HKEX filing describes prioritizing higher-value engagements while expanding standardized offerings. The issuer says reusable hardware, software modules, and system configurations can reduce engineering effort and delivery complexity. This is the company’s stated strategy and rationale, not causal proof that standardization will produce the same results for other businesses. Read the filing.
The Tool Desk
Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →The same filing reported overall gross profit margins of 30.5% in 2023, 32.3% in 2024, and 37.3% in 2025. Its adjusted total operating expenses, excluding share-based payment expenses, declined from 113.6% of revenue in 2023 to 83.4% in 2024 and 63.9% in 2025. Do not mistake those adjusted ratios for reported total expenses: 2025 total operating expenses were 107.7% of revenue, with share-based payment expenses a material factor. These issuer-specific measures also should not be treated as directly comparable to the other examples here.
Manage infrastructure investment for its full payback horizon
Capacity decisions can pressure margins before they create benefits. Consider total workload cost, quality on the intended task, latency, reliability, utilization, and deployment constraints together when comparing models or infrastructure options. No standardized apples-to-apples benchmark across vendors is established by the examples below, so vendor claims and financial disclosures need to be evaluated in their own scope.
Amazon CEO Andy Jassy said in the company’s 2025 shareholder letter that Trainium3 was 30–40% more price-performant than Trainium2 and that AWS could gain several hundred basis points of operating-margin advantage at scale compared with relying on other companies’ chips for inference. These are Amazon’s claims and expectations, not independently verified realized savings. Read the letter.
Microsoft reported a 66% Microsoft Cloud gross margin percentage in FY2026 Q3, with continuing AI infrastructure investment and growing AI product usage weighing on the measure; efficiency gains in Azure and Microsoft 365 Commercial cloud partly offset that pressure. The company also reported 20% year-over-year growth in company-wide operating income in that quarter. Those measures cover different scopes and show why Cloud gross margin and company-wide operating income should not be conflated. See Microsoft’s FY2026 Q3 results.
Recommended Free Tools
Alphabet said nearly 75% of Google Cloud customers had used its vertically optimized AI offering, and AI customers used 1.8 times as many products as customers who had not used AI. This is a company-reported association, not proof that AI use caused the greater product use. Alphabet also described infrastructure investment as increasing depreciation and data-center operating costs such as energy; depreciation rose nearly $6 billion, or 38%, from $15.3 billion in 2024 to $21.1 billion in 2025. Read the earnings-call transcript.
Best Value
Keep operating leverage from becoming indiscriminate cost cutting
Review research and development, sales and marketing, and general and administrative expense against revenue, but ask what each investment does for adoption, differentiation, and customer retention before reducing it. Expense ratios can improve as revenue scales, yet a lower ratio is not itself evidence that a business has preserved its product trajectory. Separate recurring delivery improvements from temporary timing effects and accounting adjustments, including share-based compensation where relevant.
There is no single margin target that fits every AI company, and the cited company disclosures use different scopes, definitions, and reporting periods. Survey responses and company statements also do not establish that cost changes are harmless to product growth. Treat margin work as a sequence of measured operating experiments: identify a cost driver, make a targeted change, and check both the economic result and the customer-facing guardrails before scaling it.
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.
