Alibaba claims 67% fewer tokens from a context service
Alibaba's Apsara announcements include Agent Context, which it says cuts token usage up to 67% in knowledge-heavy work. Your model bill is a retrieval problem.
The cheapest token is the one you never send. At its Apsara Conference, Alibaba laid out a full-stack AI roadmap running from chips to agents. The chip numbers will get the headlines. The line that should interest anyone paying a model bill is a context service called Agent Context, which Alibaba says cuts token usage by up to 67% in knowledge-intensive work like customer service and AI coding.
What actually happened
Alibaba announced three agent-layer pieces. AgentCore is an enterprise platform to build, run and manage agents across their lifecycle, with security at the operating layer — pitched explicitly at central governance of multi-source agents: the coding agents your developers use, the department-built ones, and third-party agent services, all under one control plane. Agent Native Cloud is the infrastructure underneath it. Agent Context is the enterprise context data service carrying the 67% claim.
The rest of the roadmap: HPN 8.0 Pro networking at 100 petabits of bandwidth across more than 130,000 ports at 800G, with failure blast radius cut from 50% to 25% versus the previous generation. A Zhenwu V900 accelerator with 216GB of memory and 1,200 GB/s inter-chip bandwidth, claiming three times its predecessor's performance, due Q1 2027. A Yitian 730 CPU with up to 40% better SPECint2017/GHz than the Yitian 710. Qwen 4 in training, with Qwen 4.5 and Qwen 5 on the roadmap scaling to 5–10 trillion parameters.
Treat the 67% as what it is: a vendor number with an "up to" in front of it, no published methodology, and no independent benchmark. We are not repeating it as a fact. We are repeating it because the direction is right even if the number is generous.
Why your token bill is a retrieval problem
Most of the AI cost conversations we walk into are about the wrong variable. Somebody wants to know whether to switch to a cheaper model. Almost nobody has looked at what they are actually sending.
You are probably paying to resend the same context repeatedly. The typical homegrown agent stuffs the whole knowledge base, or the whole conversation, into every call because that was the fastest thing to build. It works, and then the bill arrives and looks like a model problem. It is a retrieval problem wearing a model problem's clothes.
The fix is boring and it is where the money is. Retrieve the three relevant paragraphs instead of the forty-page PDF. Cache the system prompt. Summarize the conversation tail instead of replaying it. None of that requires changing vendors, and it moves the number more than a model swap usually does.
AgentCore's real pitch is agent sprawl, and that problem is already in your building. Your developers have coding agents. Marketing signed up for something. A vendor shipped an agent inside a product you already pay for. Every major cloud is now selling a governance plane for exactly that, which tells you the sprawl is universal. You do not need to buy Alibaba's — you need to be able to list what is running.
Do not buy a control plane before you have an inventory. A governance platform over an unknown fleet gives you a dashboard for agents it can see and a blind spot for the ones it cannot.
Naming note, since it will cost somebody an afternoon: AWS Bedrock already has a product called AgentCore. These are unrelated. Check which one a vendor means before you agree to anything.
Key takeaways
- Alibaba announced AgentCore, Agent Native Cloud and Agent Context at its Apsara Conference
- Agent Context is claimed to cut token usage up to 67% in knowledge-heavy scenarios — a vendor figure with no published methodology
- Also announced: HPN 8.0 Pro at 100 petabits and 130,000+ ports, Zhenwu V900 with 216GB memory due Q1 2027, Yitian 730 CPU, and Qwen 4 in training
- Context engineering usually moves your AI bill more than swapping to a cheaper model
- Retrieve narrowly, cache the system prompt, summarize the conversation tail — all vendor-independent
- Every major cloud is now selling an agent governance plane, which means agent sprawl is the universal problem
- Inventory your running agents before buying anything to govern them
- Alibaba's AgentCore is unrelated to AWS Bedrock AgentCore, despite the identical name
Before you switch models to save money, find out what you are actually sending. We audit agent context — what goes into each call, what gets resent, and what can be cached or retrieved instead — and the savings are usually in the plumbing, not the vendor. Model what your current token spend should be, or have us look at your agent stack.
Sources: Alibaba via Media OutReach newswire, HPCwire.
- #ai-agents
- #alibaba-cloud
- #token-costs
- #rag
- #agent-governance
Tommy Rush — Founder, Rush Commerce
Operator turned builder. 15+ years running operations — now shipping the systems businesses run on. More
Get The Rush Report weekly — one email, zero fluff.
Keep reading
CLOSEDQUORUM: malware that asks four AI models what to do
Cisco Talos published the first reported autonomous AI C2 implant. It polls DeepSeek, Qwen, Mistral and Gemini, then does whatever wins the vote. No operator.
Read itVerda raises $189M: another place to run inference
Helsinki's Verda hit unicorn status with a $189M Series B and a $165M revenue run rate. The neocloud tier is now real enough to quote against your hyperscaler bill.
Read it