Plan for 8x the token volume in 2027 and you can still spend what 2x would cost: $431,232 a year on a planned model mix against $432,000 on a flagship-only default. The volume is not the risk. The unplanned mix is.
2027 inference budget = Volume x Price x Mix (illustrative, Singapore list prices)
Flagship-only default Planned mix
2x volume (conservative) $432,000 $107,808
8x volume (aggressive) $1,728,000 $431,232
8x the volume on a planned mix costs about what 2x costs on the default.Budget inference as a run-rate with three dials, not as a project with one number.
In this guide
- Why AI inference cost now outgrows training, and why it keeps growing after launch
- A three-dial planning model for 2027: volume, price per unit and mix
- A worked three-scenario budget on verified Qwen list prices, with every multiplication shown
- The procurement levers that fit each part of the budget, and the cases where they do not
- An owner for every line, so variance has a name before it has a cost
Training is a project with an end date; inference is what you pay every time the model is used, for as long as it is used.
The market forecasts say the balance has already tipped. Gartner forecasts worldwide AI-optimized IaaS spending to reach $42 billion in 2026 and $66 billion in 2027, with inference at $23.3 billion in 2026 against $19 billion for training (Gartner, AI-optimized IaaS spending forecast, August 2026). That is a forecast, not a measured outcome.
Inference already leads. Gartner forecasts that 55% of AI-optimized IaaS spending in 2026 goes to inference, rising to 59% in 2027 (Gartner, AI-optimized IaaS spending forecast, August 2026).
Gartner's stated reason is the shift "from model development to production-scale deployment," and that agentic AI "amplifies compute intensity through multistep, autonomous execution, making inference the dominant consumption model." The wider budget is moving too: Gartner forecasts worldwide AI spending of $2.7 trillion in 2026, up 49.5% year over year (Gartner, worldwide AI spending forecast, September 2026).
For a CFO or CIO, the AI line in the 2027 plan is no longer mostly a build cost to amortize. It is a consumption cost that scales with adoption, like a utility.
Before you set the 2027 number, bring last quarter's token usage by workload: one minute, no login. Turn last quarter's usage into a 2027 budget →
Budget inference like a utility, because that is how it bills.
Inference grows on four axes at once, and only one of them is headcount.
1. More users. Adoption spreads from the pilot team to the function. Most budgets already model this axis.
2. More work per user. Once a tool is trusted, people send it longer documents, more turns and harder questions. Tokens per request rise even when users do not.
3. More steps per task. An agent that plans, calls tools and checks its own work makes several model calls where a chat assistant made one, and feeds each tool result back in as input. This is the multistep execution Gartner points to.
4. More workloads. Every successful use case invites the next one, usually funded from a business budget rather than IT. Boston Consulting Group reports that more than 80% of AI spending sits outside enterprise IT (BCG, Applied AI Index 2026, September 2026).
The result is that the run-rate you exit 2026 with is the floor for 2027, not the average. Cost already bites: about 20 percent of respondents report that AI-related operating costs, including token costs, constrained their AI use (McKinsey, The State of AI in 2026: On the Road to ROI, August 2026).
Treat the December run-rate as next year's floor.
An annual inference budget needs three numbers, each owned by a different person.
I call it the VPM budget: volume x price per unit x mix. It is the annual planning layer above the per-workload token forecast in our token pricing guide and the cost levers in our inference cost guide; this one is about the number finance signs.
Annual inference budget = sum over months of
Volume (input and output tokens, by workload)
x Price (list price per 1M tokens, by model ID, region and mode)
x Mix (share of volume on each model and mode)Volume comes from the business: users, adoption, new workloads and agent steps. It is the most uncertain dial, so plan it as scenarios.
Price per unit comes from the pricing page and your contract. Plan at today's list price. Prices can move in either direction, and a budget that assumes future cuts is a budget built on hope.
Mix is the dial most plans leave out. It is the share of tokens on each model tier and mode: flagship, efficient tier and batch. Mix is set by engineering and evaluation, and it can move the bill as much as volume does.
On Alibaba Cloud Model Studio, the price inputs look like this (Singapore list prices, per 1M tokens; Alibaba Cloud, Model Studio pricing and Batch inference, accessed Nov 13, 2026):
| Tier and mode | Model ID | Input | Output | Planning note |
|---|---|---|---|---|
| Flagship, real time | qwen3.8-max | $2.00 | $6.00 | Flat from 0 to 1M tokens of context |
| Efficient, real time | qwen3.8-flash | $0.15 | $0.47 | Same 1M-token context range |
| Batch | qwen-flash at 50% | $0.025 | $0.20 | List $0.05 / $0.40 for requests up to 256K input tokens; batch is 50% of real-time price on supported models |
In Singapore, batch text models are qwen3.8-max, qwen3.7-plus, qwen-max, qwen-plus and qwen-flash; qwen3.8-flash is not on the list, so a flash-tier batch line uses qwen-flash, a separate model ID that must pass your own evaluation.
Plan volume as scenarios, price at list, and mix as a decision.
Here is one company's plan, worked three ways, with every multiplication shown.
Assumptions (illustrative). December 2026 exit run-rate: 6,000M input tokens and 1,000M output tokens per month across all workloads. 2027 average monthly volume is planned at 2x (conservative), 4x (base) or 8x (aggressive, agents scale) that run-rate. Prices are the Singapore list prices above, held flat for the year. Two mixes:
Step 1: price one month of the exit run-rate.
Flagship-only default
Input 6,000M x $2.00 / 1M = $12,000
Output 1,000M x $6.00 / 1M = $6,000
Total $18,000 per month
Planned mix
qwen3.8-max 20% 1,200M x $2.00 + 200M x $6.00 = $2,400 + $1,200 = $3,600
qwen3.8-flash 60% 3,600M x $0.15 + 600M x $0.47 = $540 + $282 = $822
qwen-flash batch 20% 1,200M x $0.025 + 200M x $0.20 = $30 + $40 = $70
Total $4,492 per monthResult: the same tokens cost $18,000 or $4,492 a month. Mix alone is a 4.0x dial ($18,000 / $4,492 = 4.007).
Step 2: apply the volume scenarios for 12 months.
| 2027 scenario | Monthly volume | Flagship-only default (12 x monthly) | Planned mix (12 x monthly) |
|---|---|---|---|
| Conservative, 2x | 12,000M in / 2,000M out | 12 x $36,000 = $432,000 | 12 x $8,984 = $107,808 |
| Base, 4x | 24,000M in / 4,000M out | 12 x $72,000 = $864,000 | 12 x $17,968 = $215,616 |
| Aggressive, 8x | 48,000M in / 8,000M out | 12 x $144,000 = $1,728,000 | 12 x $35,936 = $431,232 |
Step 3: add a labelled contingency. Retries, long-tail requests and unplanned workloads are real. Assume a 15% buffer on the planned mix:
Base case, planned mix $215,616 x 1.15 = $247,958.40
Aggressive case, planned mix $431,232 x 1.15 = $495,916.80Result: the volume range (2x to 8x) is a 4x spread. The mix choice is also a 4.0x spread. The aggressive scenario on a planned mix ($431,232) lands within $768 of the conservative scenario on the default ($432,000).
You cannot control 2027 volume precisely, but you can decide the mix now.
Find the dial that moves your 2027 bill most
Send monthly token volumes by workload and the models you call today. The team will work through volume, price and mix with you at Model Studio list prices, and show which dial moves your 2027 number most. Price my 2027 inference scenarios →
Decide the mix before you argue about the volume.
Each lever fits one shape of demand, and some of them cannot be combined.
| Lever | Fits | How it bills | Watch out for |
|---|---|---|---|
| Pay as you go | New, variable or unproven workloads | Per token at list price | No commitment, no protection from your own growth |
| Savings plans | A steady, forecastable base of token spend | Model Studio says "you can use savings plans to offset the cost" of token billing | Batch is covered only by the AI Universal Savings Plan, not other prepaid Savings Plans |
| Throughput Reservation | Predictable peak traffic that cannot tolerate rate limits | Prepaid dedicated tokens-per-minute capacity for a specified model, in Beijing and Singapore | Overflow either falls back to pay as you go or returns errors, depending on the policy you choose |
| Batch | Work that can wait: reports, enrichment, evaluation, document review | 50% of real-time price on supported models | Eligible for the AI Universal Savings Plan, but not other prepaid Savings Plans, free quota or context caching |
| Coding Plan | AI coding tools | Fixed monthly fee with a monthly request quota | Scoped to coding tools, not general workloads |
Sources: Alibaba Cloud, What is Model Studio, Throughput Reservation and Batch inference, accessed Nov 13, 2026. Check prices for savings plans and Throughput Reservation in the console; they depend on model, region and term.
The pattern for a CFO: commit to the floor, stay flexible on the growth. Put the December run-rate's steady base on a savings plan, reserve throughput only where an outage at peak would cost more than the reservation, move everything that can wait to batch, and keep new workloads on pay as you go until they have three months of history.
When not to commit. Do not sign a commitment sized to the aggressive scenario; size it to the floor you are confident in. Do not reserve throughput for a workload that has not run in production long enough to show its peak.
Still in pilot? Skip this model. With less than a quarter of real logs, volume is a guess and the VPM budget only dresses it up. Budget pilots as a fixed experiment cap instead.
If the volume is large and steady enough that self-hosting open weights looks attractive, weigh it the same way we weighed the build-or-partner call in Build vs. Buy, or Both? The Enterprise AI Decision.
Commit to the floor, keep the growth on demand.
A budget with three dials needs three owners, plus one person who reconciles them.
| Budget line | Owner | What they commit to | Monthly signal |
|---|---|---|---|
| Volume | Business owner of each workload | Users, adoption and new workloads per scenario | Tokens by workload vs plan |
| Price | Procurement and FinOps | List prices, savings plans, reservations | Effective price per 1M tokens vs list |
| Mix | AI platform or engineering lead | Share of tokens per tier and mode, gated by evaluation | Share on flagship vs plan |
| Variance | FP&A, with the CIO | Monthly reforecast and which dial moved | Actual vs scenario, by dial |
Ownership matters because the spend is spread across the business. With more than 80% of AI spending outside enterprise IT (BCG, Applied AI Index 2026, September 2026), a central IT budget cannot see most of the volume dial on its own.
Three controls make the owners real:
Every dial gets an owner, or the dial owns you.
You can produce a board-ready VPM budget in four weeks.
Plan the run-rate in three dials, and commit only to what you can see.
Key takeaways
- Inference is the run-rate. Gartner forecasts inference at 55% of AI-optimized IaaS spending in 2026, rising to 59% in 2027.
- Mix is as big a dial as volume. In the worked plan, mix moved one month from $18,000 to $4,492, and 8x volume on a planned mix ($431,232) cost about what 2x cost on the default ($432,000).
- Commit to the floor. Savings plans for the steady base, Throughput Reservation for proven peaks, batch for work that can wait, pay as you go for the rest.
Bring a budget finance can sign, not a guess
Share your workloads, owners and last quarter's token usage in a short form. The team will review your volume scenarios, mix and commitment sizing with you, and flag which procurement lever fits each workload. Get my 2027 VPM budget reviewed →
See more of the enterprise AI portfolio at qwen.ai. Lasting returns come from treating AI like infrastructure: planned over years, delivered in stages, with Qwen alongside you so the early choices hold up and the value keeps growing.
About Qwen. Qwen gives enterprises an open route into production AI: one model family covering language, vision, audio, coding, embeddings, and agents, as open weights and managed APIs. Built on curated data and ongoing research, Qwen supports creating, testing, and running AI systems for decisions that matter, from self-serve developer access to dedicated inference.
Sources: Gartner, Gartner Forecasts Worldwide AI-Optimized IaaS Spending to Grow 96% Through 2026 (August 2026); Gartner, Gartner Forecasts Worldwide AI Spending to Grow 49.5% in 2026 (September 2026); BCG, Applied AI Index 2026 (September 2026); McKinsey, The State of AI in 2026: On the Road to ROI (August 2026); Alibaba Cloud, Model Studio pricing (accessed Nov 13, 2026); Alibaba Cloud, Batch inference (accessed Nov 13, 2026); Alibaba Cloud, Throughput Reservation (accessed Nov 13, 2026); Alibaba Cloud, What is Model Studio (accessed Nov 13, 2026).
Johnny Mai - Director, Gen AI Product Strategy & GTM, Qwen.
Disclaimer: This post is provided for general information only. Worked examples are illustrative and use stated assumptions; your costs will vary. Prices are Alibaba Cloud Model Studio list prices for the Singapore region as published in November 2026, vary by region, and may change over time. Figures from third-party research (Gartner 2026; BCG 2026; McKinsey 2026) are quoted as published in those reports and are not independently verified by Qwen or Alibaba Cloud. Product and service availability, features, and program terms vary by region and may change over time. Nothing in this post constitutes professional, legal, or investment advice. © 2026 Qwen. All rights reserved.
7 Lessons from Taking Enterprise AI from Pilot to Production
Batch vs. Real-Time Inference: Which Workloads Should Wait for Half Price
21 posts | 2 followers
FollowJohnny Mai - October 7, 2026
Johnny Mai - September 10, 2026
Johnny Mai - November 13, 2026
Johnny Mai - November 9, 2026
Alibaba Cloud Community - September 14, 2026
Alibaba Cloud Community - August 5, 2026
21 posts | 2 followers
Follow
DevOps Solution
Accelerate software development and delivery by integrating DevOps with the cloud
Learn More
API Gateway
API Gateway provides you with high-performance and high-availability API hosting services to deploy and release your APIs on Alibaba Cloud products.
Learn More
Architecture and Structure Design
Customized infrastructure to ensure high availability, scalability and high-performance
Learn More
OpenAPI Explorer
OpenAPI Explorer allows you to call an API through its web interface or WebCLI, and view the entire process.
Learn MoreMore Posts by Johnny Mai